Spline2DSpec.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_GPUCOMMON_TPCFASTTRANSFORMATION_SPLINE2DSPEC_H
#define ALICEO2_GPUCOMMON_TPCFASTTRANSFORMATION_SPLINE2DSPEC_H
 
#include "Spline1D.h"
#include "FlatObject.h"
#include "GPUCommonDef.h"
#include "SplineUtil.h"
 
#if !defined(__CLING__) && !defined(G__ROOT) && !defined(GPUCA_GPUCODE) && !defined(GPUCA_NO_VC)
#include <Vc/Vc>
#include <Vc/SimdArray>
#endif
 
class TFile;
 
namespace o2
{
namespace gpu
{
 
template <typename DataT>
class Spline2DContainer : public FlatObject
{
 public:
  typedef typename Spline1D<DataT>::SafetyLevel SafetyLevel;
  typedef typename Spline1D<DataT>::Knot Knot;
 
 
  GPUd() static constexpr int32_t getVersion() { return (1 << 16) + Spline1D<DataT>::getVersion(); }
 
 
  Spline2DContainer() = default;
 
  Spline2DContainer(const Spline2DContainer&) = delete;
 
  ~Spline2DContainer() = default;
 
 
#if !defined(GPUCA_GPUCODE) && !defined(GPUCA_STANDALONE)
  void approximateFunction(double x1Min, double x1Max, double x2Min, double x2Max,
                           std::function<void(double x1, double x2, double f[/*mYdim*/])> F,
                           int32_t nAuxiliaryDataPointsU1 = 4, int32_t nAuxiliaryDataPointsU2 = 4);
 
  void approximateFunctionViaDataPoints(double x1Min, double x1Max, double x2Min, double x2Max,
                                        std::function<void(double x1, double x2, double f[])> F,
                                        int32_t nAuxiliaryDataPointsX1, int32_t nAuxiliaryDataPointsX2);
#endif
 
 
#if !defined(GPUCA_GPUCODE) && !defined(GPUCA_STANDALONE)
  int32_t writeToFile(TFile& outf, const char* name);
 
  static Spline2DContainer* readFromFile(TFile& inpf, const char* name);
#endif
 
 
  GPUd() int32_t getYdimensions() const { return mYdim; }
 
  GPUd() static constexpr size_t getParameterAlignmentBytes() { return 16; }
 
  GPUd() int32_t getNumberOfParameters() const { return this->calcNumberOfParameters(mYdim); }
 
  GPUd() size_t getSizeOfParameters() const { return sizeof(DataT) * this->getNumberOfParameters(); }
 
  GPUd() int32_t getNumberOfKnots() const { return mGridX1.getNumberOfKnots() * mGridX2.getNumberOfKnots(); }
 
  GPUd() const Spline1D<DataT>& getGridX1() const { return mGridX1; }
 
  GPUd() const Spline1D<DataT>& getGridX2() const { return mGridX2; }
 
  GPUd() const Spline1D<DataT>& getGrid(int32_t ix) const { return (ix == 0) ? mGridX1 : mGridX2; }
 
  GPUd() void getKnotU(int32_t iKnot, int32_t& u1, int32_t& u2) const
  {
    u1 = mGridX1.getKnot(iKnot % mGridX1.getNumberOfKnots()).getU();
    u2 = mGridX2.getKnot(iKnot / mGridX1.getNumberOfKnots()).getU();
  }
 
  GPUd() int32_t getKnotIndex(int32_t iKnotX1, int32_t iKnotX2) const
  {
    return mGridX1.getNumberOfKnots() * iKnotX2 + iKnotX1;
  }
 
  GPUd() DataT* getParameters() { return mParameters; }
 
  GPUd() const DataT* getParameters() const { return mParameters; }
 
 
  GPUd() size_t getGridX1Offset() const { return mGridX1.getFlatBufferPtr() - mFlatBufferPtr; }
 
  GPUd() size_t getGridX2Offset() const { return mGridX2.getFlatBufferPtr() - mFlatBufferPtr; }
 
  GPUd() void setXrange(DataT x1Min, DataT x1Max, DataT x2Min, DataT x2Max)
  {
    mGridX1.setXrange(x1Min, x1Max);
    mGridX2.setXrange(x2Min, x2Max);
  }
 
  void print() const;
 
 
  GPUd() int32_t calcNumberOfParameters(int32_t nYdim) const { return (4 * nYdim) * getNumberOfKnots(); }
 
 
#if !defined(GPUCA_GPUCODE) && !defined(GPUCA_STANDALONE) // code invisible on GPU and in the standalone compilation
  static int32_t test(const bool draw = 0, const bool drawDataPoints = 1);
#endif
 
 
  using FlatObject::getBufferAlignmentBytes;
  using FlatObject::getClassAlignmentBytes;
 
#if !defined(GPUCA_GPUCODE)
  void cloneFromObject(const Spline2DContainer& obj, char* newFlatBufferPtr);
  void moveBufferTo(char* newBufferPtr);
#endif
 
  using FlatObject::releaseInternalBuffer;
 
  void destroy();
  void setActualBufferAddress(char* actualFlatBufferPtr);
  void setFutureBufferAddress(char* futureFlatBufferPtr);
 
 protected:
#if !defined(GPUCA_GPUCODE)
  void recreate(int32_t nYdim, int32_t nKnotsX1, int32_t nKnotsX2);
 
  void recreate(int32_t nYdim, int32_t nKnotsX1, const int32_t knotU1[], int32_t nKnotsX2, const int32_t knotU2[]);
#endif
 
 
  int32_t mYdim = 0;            
  Spline1D<DataT> mGridX1;      
  Spline1D<DataT> mGridX2;      
  DataT* mParameters = nullptr; 
 
  ClassDefNV(Spline2DContainer, 1);
};
 
template <typename DataT, int32_t YdimT, int32_t SpecT>
class Spline2DSpec;
 
template <typename DataT, int32_t YdimT>
class Spline2DSpec<DataT, YdimT, 0>
  : public Spline2DContainer<DataT>
{
  typedef Spline2DContainer<DataT> TBase;
 
 public:
  typedef typename TBase::SafetyLevel SafetyLevel;
  typedef typename TBase::Knot Knot;
 
 
  GPUd() void interpolate(DataT x1, DataT x2, GPUgeneric() DataT S[/*mYdim*/]) const
  {
    interpolateU<SafetyLevel::kSafe>(mYdim, mParameters, mGridX1.convXtoU(x1), mGridX2.convXtoU(x2), S);
  }
 
  template <SafetyLevel SafeT = SafetyLevel::kSafe>
  GPUd() void interpolateUold(int32_t inpYdim, GPUgeneric() const DataT Parameters[],
                              DataT u1, DataT u2, GPUgeneric() DataT S[/*inpYdim*/]) const
  {
 
    const auto nYdimTmp = SplineUtil::getNdim<YdimT>(inpYdim);
    const int32_t nYdim = nYdimTmp.get();
 
    const auto maxYdim = SplineUtil::getMaxNdim<YdimT>(inpYdim);
    const int32_t maxYdim4 = 4 * maxYdim.get();
 
    const auto nYdim2 = nYdim * 2;
    const auto nYdim4 = nYdim * 4;
 
    const DataT& u = u1;
    const DataT& v = u2;
    int32_t nu = mGridX1.getNumberOfKnots();
    int32_t iu = mGridX1.template getLeftKnotIndexForU<SafeT>(u);
    int32_t iv = mGridX2.template getLeftKnotIndexForU<SafeT>(v);
 
    const typename TBase::Knot& knotU = mGridX1.template getKnot<SafetyLevel::kNotSafe>(iu);
    const typename TBase::Knot& knotV = mGridX2.template getKnot<SafetyLevel::kNotSafe>(iv);
 
    const DataT* par00 = Parameters + (nu * iv + iu) * nYdim4; // values { {Y1,Y2,Y3}, {Y1,Y2,Y3}'v, {Y1,Y2,Y3}'u, {Y1,Y2,Y3}''vu } at {u0, v0}
    const DataT* par10 = par00 + nYdim4;                       // values { ... } at {u1, v0}
    const DataT* par01 = par00 + nYdim4 * nu;                  // values { ... } at {u0, v1}
    const DataT* par11 = par01 + nYdim4;                       // values { ... } at {u1, v1}
 
    DataT Su0[maxYdim4]; // values { {Y1,Y2,Y3,Y1'v,Y2'v,Y3'v}(v0), {Y1,Y2,Y3,Y1'v,Y2'v,Y3'v}(v1) }, at u0
    DataT Du0[maxYdim4]; // derivatives {}'_u  at u0
    DataT Su1[maxYdim4]; // values { {Y1,Y2,Y3,Y1'v,Y2'v,Y3'v}(v0), {Y1,Y2,Y3,Y1'v,Y2'v,Y3'v}(v1) }, at u1
    DataT Du1[maxYdim4]; // derivatives {}'_u  at u1
 
    for (int32_t i = 0; i < nYdim2; i++) {
      Su0[i] = par00[i];
      Su0[nYdim2 + i] = par01[i];
 
      Du0[i] = par00[nYdim2 + i];
      Du0[nYdim2 + i] = par01[nYdim2 + i];
 
      Su1[i] = par10[i];
      Su1[nYdim2 + i] = par11[i];
 
      Du1[i] = par10[nYdim2 + i];
      Du1[nYdim2 + i] = par11[nYdim2 + i];
    }
 
    DataT parU[maxYdim4]; // interpolated values { {Y1,Y2,Y3,Y1'v,Y2'v,Y3'v}(v0), {Y1,Y2,Y3,Y1'v,Y2'v,Y3'v}(v1) } at u
 
    typedef Spline1DSpec<DataT, 4 * YdimT, 0> TGridX1;
    const TGridX1& gridX1 = reinterpret_cast<const TGridX1&>(mGridX1);
 
    gridX1.interpolateU(nYdim4, knotU, Su0, Du0, Su1, Du1, u, parU);
 
    const DataT* Sv0 = parU + 0;
    const DataT* Dv0 = parU + nYdim;
    const DataT* Sv1 = parU + nYdim2;
    const DataT* Dv1 = parU + nYdim2 + nYdim;
 
    typedef Spline1DSpec<DataT, YdimT, 0> TGridX2;
    const TGridX2& gridX2 = reinterpret_cast<const TGridX2&>(mGridX2);
    gridX2.interpolateU(nYdim, knotV, Sv0, Dv0, Sv1, Dv1, v, S);
  }
 
  template <SafetyLevel SafeT = SafetyLevel::kSafe>
  GPUd() void interpolateU(int32_t inpYdim, GPUgeneric() const DataT Parameters[],
                           DataT u1, DataT u2, GPUgeneric() DataT S[/*inpYdim*/]) const
  {
 
    const auto nYdimTmp = SplineUtil::getNdim<YdimT>(inpYdim);
    const int32_t nYdim = nYdimTmp.get();
 
    // const auto maxYdim = SplineUtil::getMaxNdim<YdimT>(inpYdim);
    // const int32_t maxYdim4 = 4 * maxYdim.get();
 
    // const auto nYdim2 = nYdim * 2;
    const auto nYdim4 = nYdim * 4;
 
    const DataT& u = u1;
    const DataT& v = u2;
    int32_t nu = mGridX1.getNumberOfKnots();
    int32_t iu = mGridX1.template getLeftKnotIndexForU<SafeT>(u);
    int32_t iv = mGridX2.template getLeftKnotIndexForU<SafeT>(v);
 
    const typename TBase::Knot& knotU = mGridX1.template getKnot<SafetyLevel::kNotSafe>(iu);
    const typename TBase::Knot& knotV = mGridX2.template getKnot<SafetyLevel::kNotSafe>(iv);
 
    const DataT* A = Parameters + (nu * iv + iu) * nYdim4; // values { {Y1,Y2,Y3}, {Y1,Y2,Y3}'v, {Y1,Y2,Y3}'u, {Y1,Y2,Y3}''vu } at {u0, v0}
    const DataT* B = A + nYdim4 * nu;                      // values { ... } at {u0, v1}
 
    DataT dSl, dDl, dSr, dDr;
    mGridX1.getUderivatives(knotU, u, dSl, dDl, dSr, dDr);
    DataT dSd, dDd, dSu, dDu;
    mGridX2.getUderivatives(knotV, v, dSd, dDd, dSu, dDu);
 
    // when nYdim == 1:
    // S = dSl * (dSd * A[0] + dDd * A[1]) + dDl * (dSd * A[2] + dDd * A[3]) +
    //     dSr * (dSd * A[4] + dDd * A[5]) + dDr * (dSd * A[6] + dDd * A[7]) +
    //     dSl * (dSu * B[0] + dDu * B[1]) + dDl * (dSu * B[2] + dDu * B[3]) +
    //     dSr * (dSu * B[4] + dDu * B[5]) + dDr * (dSu * B[6] + dDu * B[7]);
 
    DataT a[8] = {dSl * dSd, dSl * dDd, dDl * dSd, dDl * dDd,
                  dSr * dSd, dSr * dDd, dDr * dSd, dDr * dDd};
    DataT b[8] = {dSl * dSu, dSl * dDu, dDl * dSu, dDl * dDu,
                  dSr * dSu, dSr * dDu, dDr * dSu, dDr * dDu};
 
    // S = sum a[i]*A[i] + b[i]*B[i]
 
    for (int32_t dim = 0; dim < nYdim; dim++) {
      S[dim] = 0;
      for (int32_t i = 0; i < 8; i++) {
        S[dim] += a[i] * A[nYdim * i + dim] + b[i] * B[nYdim * i + dim];
      }
    }
  }
 
 protected:
  using TBase::mGridX1;
  using TBase::mGridX2;
  using TBase::mParameters;
  using TBase::mYdim;
  using TBase::TBase; // inherit constructors and hide them
};
 
template <typename DataT, int32_t YdimT>
class Spline2DSpec<DataT, YdimT, 1>
  : public Spline2DSpec<DataT, YdimT, 0>
{
  typedef Spline2DContainer<DataT> TVeryBase;
  typedef Spline2DSpec<DataT, YdimT, 0> TBase;
 
 public:
  typedef typename TVeryBase::SafetyLevel SafetyLevel;
 
#if !defined(GPUCA_GPUCODE)
  Spline2DSpec() : Spline2DSpec(2, 2) {}
 
  Spline2DSpec(int32_t nKnotsX1, int32_t nKnotsX2) : TBase()
  {
    recreate(nKnotsX1, nKnotsX2);
  }
  Spline2DSpec(int32_t nKnotsX1, const int32_t knotU1[],
               int32_t nKnotsX2, const int32_t knotU2[])
    : TBase()
  {
    recreate(nKnotsX1, knotU1, nKnotsX2, knotU2);
  }
  Spline2DSpec(const Spline2DSpec& v) : TBase()
  {
    TBase::cloneFromObject(v, nullptr);
  }
  void recreate(int32_t nKnotsX1, int32_t nKnotsX2)
  {
    TBase::recreate(YdimT, nKnotsX1, nKnotsX2);
  }
 
  void recreate(int32_t nKnotsX1, const int32_t knotU1[],
                int32_t nKnotsX2, const int32_t knotU2[])
  {
    TBase::recreate(YdimT, nKnotsX1, knotU1, nKnotsX2, knotU2);
  }
#endif
 
  GPUd() constexpr int32_t getYdimensions() const { return YdimT; }
 
  GPUd() int32_t getNumberOfParameters() const { return (4 * YdimT) * getNumberOfKnots(); }
 
  GPUd() size_t getSizeOfParameters() const { return (sizeof(DataT) * 4 * YdimT) * getNumberOfKnots(); }
 
 
  template <SafetyLevel SafeT = SafetyLevel::kSafe>
  GPUd() void interpolateU(GPUgeneric() const DataT Parameters[],
                           DataT u1, DataT u2, GPUgeneric() DataT S[/*nYdim*/]) const
  {
    TBase::template interpolateU<SafeT>(YdimT, Parameters, u1, u2, S);
  }
 
  template <SafetyLevel SafeT = SafetyLevel::kSafe>
  GPUd() void interpolateUold(GPUgeneric() const DataT Parameters[],
                              DataT u1, DataT u2, GPUgeneric() DataT S[/*nYdim*/]) const
  {
    TBase::template interpolateUold<SafeT>(YdimT, Parameters, u1, u2, S);
  }
 
  using TBase::getNumberOfKnots;
 
 private:
#if !defined(GPUCA_GPUCODE)
  using TBase::recreate;
#endif
  using TBase::interpolateU;
};
 
template <typename DataT, int32_t YdimT>
class Spline2DSpec<DataT, YdimT, 2>
  : public Spline2DSpec<DataT, YdimT, 0>
{
  typedef Spline2DContainer<DataT> TVeryBase;
  typedef Spline2DSpec<DataT, YdimT, 0> TBase;
 
 public:
  typedef typename TVeryBase::SafetyLevel SafetyLevel;
 
#if !defined(GPUCA_GPUCODE)
  Spline2DSpec() : Spline2DSpec(0, 2, 2) {}
 
  Spline2DSpec(int32_t nYdim, int32_t nKnotsX1, int32_t nKnotsX2) : TBase()
  {
    TBase::recreate(nYdim, nKnotsX1, nKnotsX2);
  }
 
  Spline2DSpec(int32_t nYdim, int32_t nKnotsX1, const int32_t knotU1[],
               int32_t nKnotsX2, const int32_t knotU2[]) : TBase()
  {
    TBase::recreate(nYdim, nKnotsX1, knotU1, nKnotsX2, knotU2);
  }
 
  Spline2DSpec(const Spline2DSpec& v) : TBase()
  {
    cloneFromObject(v, nullptr);
  }
 
  void recreate(int32_t nYdim, int32_t nKnotsX1, int32_t nKnotsX2)
  {
    TBase::recreate(nYdim, nKnotsX1, nKnotsX2);
  }
 
  void recreate(int32_t nYdim, int32_t nKnotsX1, const int32_t knotU1[],
                int32_t nKnotsX2, const int32_t knotU2[])
  {
    TBase::recreate(nYdim, nKnotsX1, knotU1, nKnotsX2, knotU2);
  }
#endif
 
 
  using TBase::interpolateU;
};
 
template <typename DataT>
class Spline2DSpec<DataT, 1, 3>
  : public Spline2DSpec<DataT, 1, SplineUtil::getSpec(999)>
{
  typedef Spline2DSpec<DataT, 1, SplineUtil::getSpec(999)> TBase;
 
 public:
  using TBase::TBase; // inherit constructors
 
  GPUd() DataT interpolate(DataT x1, DataT x2) const
  {
    DataT S = 0.;
    TBase::interpolate(x1, x2, &S);
    return S;
  }
 
  // this parent method should be public anyhow,
  // but w/o this extra declaration compiler gets confused
  using TBase::interpolate;
};
} // namespace gpu
} // namespace o2
 
#endif