Skip to content

Commit

Permalink
cp
Browse files Browse the repository at this point in the history
  • Loading branch information
@mconcas
mconcas committed Oct 26, 2024
1 parent 7393473 commit 5b6577c
Show file tree
Hide file tree
Showing 11 changed files with 427 additions and 214 deletions.
37 changes: 37 additions & 0 deletions Common/MathUtils/include/MathUtils/SMatrixGPU.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -446,6 +446,8 @@ class SMatrixGPU
GPUdi() SMatrixGPU(SMatrixNoInit) {}
GPUd() SMatrixGPU(SMatrixIdentity);
GPUd() SMatrixGPU(const SMatrixGPU<T, D1, D2, R>& rhs);
template <class R2>
GPUd() SMatrixGPU(const SMatrixGPU<T, D1, D2, R2>& rhs);
template <class A, class R2>
GPUd() SMatrixGPU(const Expr<A, T, D1, D2, R2>& rhs);
template <class M>
Expand Down Expand Up @@ -497,6 +499,11 @@ class SMatrixGPU
GPUd() SMatrixRowGPU operator[](unsigned int i) { return SMatrixRowGPU(*this, i); }
template <class R2>
GPUd() SMatrixGPU<T, D1, D2, R>& operator+=(const SMatrixGPU<T, D1, D2, R2>& rhs);
GPUd() SMatrixGPU<T, D1, D2, R>& operator*=(const T& rhs);
template <class R2>
GPUd() SMatrixGPU<T, D1, D2, R>& operator*=(const SMatrixGPU<T, D1, D2, R2>& rhs);
template <class A, class R2>
GPUd() SMatrixGPU<T, D1, D2, R>& operator*=(const Expr<A, T, D1, D2, R2>& rhs);

GPUd() bool Invert();
GPUd() bool IsInUse(const T* p) const;
Expand Down Expand Up @@ -528,6 +535,13 @@ GPUdi() SMatrixGPU<T, D1, D2, R>::SMatrixGPU(const SMatrixGPU<T, D1, D2, R>& rhs
mRep = rhs.mRep;
}

template <class T, unsigned int D1, unsigned int D2, class R>
template <class R2>
GPUd() SMatrixGPU<T, D1, D2, R>::SMatrixGPU(const SMatrixGPU<T, D1, D2, R2>& rhs)
{
operator=(rhs);
}

template <class T, unsigned int D1, unsigned int D2, class R>
GPUdi() T* SMatrixGPU<T, D1, D2, R>::begin()
{
Expand Down Expand Up @@ -1387,6 +1401,29 @@ GPUdi() SMatrixGPU<T, D1, D2, R>& SMatrixGPU<T, D1, D2, R>::operator+=(const SMa
return *this;
}

template <class T, unsigned int D1, unsigned int D2, class R>
GPUdi() SMatrixGPU<T, D1, D2, R>& SMatrixGPU<T, D1, D2, R>::operator*=(const T & rhs)
{
for (unsigned int i = 0; i < R::kSize; ++i) {
mRep.Array()[i] *= rhs;
}
return *this;
}

template <class T, unsigned int D1, unsigned int D2, class R>
template <class R2>
GPUdi() SMatrixGPU<T, D1, D2, R>& SMatrixGPU<T, D1, D2, R>::operator*=(const SMatrixGPU<T, D1, D2, R2>& rhs)
{
return operator=(*this* rhs);
}

template <class T, unsigned int D1, unsigned int D2, class R>
template <class A, class R2>
GPUdi() SMatrixGPU<T, D1, D2, R>& SMatrixGPU<T, D1, D2, R>::operator*=(const Expr<A, T, D1, D2, R2>& rhs)
{
return operator=(*this* rhs);
}

template <class T, unsigned int D1, unsigned int D2, class R>
struct TranspPolicyGPU {
enum {
Expand Down
13 changes: 7 additions & 6 deletions DataFormats/Reconstruction/include/ReconstructionDataFormats/TrackParametrizationWithError.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,7 @@
#define INCLUDE_RECONSTRUCTIONDATAFORMATS_TRACKPARAMETRIZATIONWITHERROR_H_

#include "ReconstructionDataFormats/TrackParametrization.h"
#include <MathUtils/Cartesian.h>

namespace o2
{
Expand All @@ -38,8 +39,8 @@ class TrackParametrizationWithError : public TrackParametrization<value_T>
#endif

using covMat_t = gpu::gpustd::array<value_t, kCovMatSize>;
using MatrixDSym5 = ROOT::Math::SMatrix<double, kNParams, kNParams, ROOT::Math::MatRepSym<double, kNParams>>;
using MatrixD5 = ROOT::Math::SMatrix<double, kNParams, kNParams, ROOT::Math::MatRepStd<double, kNParams, kNParams>>;
using MatrixDSym5 = o2::math_utils::SMatrix<double, kNParams, kNParams, o2::math_utils::MatRepSym<double, kNParams>>;
using MatrixD5 = o2::math_utils::SMatrix<double, kNParams, kNParams, o2::math_utils::MatRepStd<double, kNParams, kNParams>>;

GPUd() TrackParametrizationWithError();
GPUd() TrackParametrizationWithError(value_t x, value_t alpha, const params_t& par, const covMat_t& cov, int charge = 1, const PID pid = PID::Pion);
Expand Down Expand Up @@ -100,12 +101,12 @@ class TrackParametrizationWithError : public TrackParametrization<value_T>
template <typename T>
GPUd() value_t getPredictedChi2(const BaseCluster<T>& p) const;

void buildCombinedCovMatrix(const TrackParametrizationWithError& rhs, MatrixDSym5& cov) const;
value_t getPredictedChi2(const TrackParametrizationWithError& rhs, MatrixDSym5& covToSet) const;
GPUd() void buildCombinedCovMatrix(const TrackParametrizationWithError& rhs, MatrixDSym5& cov) const;
GPUd() value_t getPredictedChi2(const TrackParametrizationWithError& rhs, MatrixDSym5& covToSet) const;
GPUd() value_t getPredictedChi2(const TrackParametrizationWithError& rhs) const;
GPUd() value_t getPredictedChi2Quiet(const TrackParametrizationWithError& rhs) const;
bool update(const TrackParametrizationWithError& rhs, const MatrixDSym5& covInv);
bool update(const TrackParametrizationWithError& rhs);
GPUd() bool update(const TrackParametrizationWithError& rhs, const MatrixDSym5& covInv);
GPUd() bool update(const TrackParametrizationWithError& rhs);

GPUd() bool update(const dim2_t& p, const dim3_t& cov);
GPUd() bool update(const value_t* p, const value_t* cov);
Expand Down
52 changes: 25 additions & 27 deletions DataFormats/Reconstruction/src/TrackParametrizationWithError.cxx
View file
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,7 @@
#include "ReconstructionDataFormats/Vertex.h"
#include "ReconstructionDataFormats/DCA.h"
#include <GPUCommonLogger.h>
#include <cfloat>

#ifndef GPUCA_GPUCODE_DEVICE
#include <iostream>
Expand Down Expand Up @@ -754,30 +755,6 @@ GPUd() auto TrackParametrizationWithError<value_T>::getPredictedChi2Quiet(const
return (d * (szz * d - sdz * z) + z * (sdd * z - d * sdz)) / det;
}

#if !defined(GPUCA_GPUCODE) && !defined(GPUCA_STANDALONE) // Disable function relying on ROOT SMatrix on GPU

//______________________________________________
template <typename value_T>
void TrackParametrizationWithError<value_T>::buildCombinedCovMatrix(const TrackParametrizationWithError<value_T>& rhs, MatrixDSym5& cov) const
{
// fill combined cov.matrix (NOT inverted)
cov(kY, kY) = static_cast<double>(getSigmaY2()) + static_cast<double>(rhs.getSigmaY2());
cov(kZ, kY) = static_cast<double>(getSigmaZY()) + static_cast<double>(rhs.getSigmaZY());
cov(kZ, kZ) = static_cast<double>(getSigmaZ2()) + static_cast<double>(rhs.getSigmaZ2());
cov(kSnp, kY) = static_cast<double>(getSigmaSnpY()) + static_cast<double>(rhs.getSigmaSnpY());
cov(kSnp, kZ) = static_cast<double>(getSigmaSnpZ()) + static_cast<double>(rhs.getSigmaSnpZ());
cov(kSnp, kSnp) = static_cast<double>(getSigmaSnp2()) + static_cast<double>(rhs.getSigmaSnp2());
cov(kTgl, kY) = static_cast<double>(getSigmaTglY()) + static_cast<double>(rhs.getSigmaTglY());
cov(kTgl, kZ) = static_cast<double>(getSigmaTglZ()) + static_cast<double>(rhs.getSigmaTglZ());
cov(kTgl, kSnp) = static_cast<double>(getSigmaTglSnp()) + static_cast<double>(rhs.getSigmaTglSnp());
cov(kTgl, kTgl) = static_cast<double>(getSigmaTgl2()) + static_cast<double>(rhs.getSigmaTgl2());
cov(kQ2Pt, kY) = static_cast<double>(getSigma1PtY()) + static_cast<double>(rhs.getSigma1PtY());
cov(kQ2Pt, kZ) = static_cast<double>(getSigma1PtZ()) + static_cast<double>(rhs.getSigma1PtZ());
cov(kQ2Pt, kSnp) = static_cast<double>(getSigma1PtSnp()) + static_cast<double>(rhs.getSigma1PtSnp());
cov(kQ2Pt, kTgl) = static_cast<double>(getSigma1PtTgl()) + static_cast<double>(rhs.getSigma1PtTgl());
cov(kQ2Pt, kQ2Pt) = static_cast<double>(getSigma1Pt2()) + static_cast<double>(rhs.getSigma1Pt2());
}

//______________________________________________
template <typename value_T>
GPUd() auto TrackParametrizationWithError<value_T>::getPredictedChi2(const TrackParametrizationWithError<value_T>& rhs) const -> value_t
Expand Down Expand Up @@ -819,6 +796,28 @@ GPUd() auto TrackParametrizationWithError<value_T>::getPredictedChi2(const Track
return chi2diag + 2. * chi2ndiag;
}

//______________________________________________
template <typename value_T>
GPUd() void TrackParametrizationWithError<value_T>::buildCombinedCovMatrix(const TrackParametrizationWithError<value_T>& rhs, MatrixDSym5& cov) const
{
// fill combined cov.matrix (NOT inverted)
cov(kY, kY) = static_cast<double>(getSigmaY2()) + static_cast<double>(rhs.getSigmaY2());
cov(kZ, kY) = static_cast<double>(getSigmaZY()) + static_cast<double>(rhs.getSigmaZY());
cov(kZ, kZ) = static_cast<double>(getSigmaZ2()) + static_cast<double>(rhs.getSigmaZ2());
cov(kSnp, kY) = static_cast<double>(getSigmaSnpY()) + static_cast<double>(rhs.getSigmaSnpY());
cov(kSnp, kZ) = static_cast<double>(getSigmaSnpZ()) + static_cast<double>(rhs.getSigmaSnpZ());
cov(kSnp, kSnp) = static_cast<double>(getSigmaSnp2()) + static_cast<double>(rhs.getSigmaSnp2());
cov(kTgl, kY) = static_cast<double>(getSigmaTglY()) + static_cast<double>(rhs.getSigmaTglY());
cov(kTgl, kZ) = static_cast<double>(getSigmaTglZ()) + static_cast<double>(rhs.getSigmaTglZ());
cov(kTgl, kSnp) = static_cast<double>(getSigmaTglSnp()) + static_cast<double>(rhs.getSigmaTglSnp());
cov(kTgl, kTgl) = static_cast<double>(getSigmaTgl2()) + static_cast<double>(rhs.getSigmaTgl2());
cov(kQ2Pt, kY) = static_cast<double>(getSigma1PtY()) + static_cast<double>(rhs.getSigma1PtY());
cov(kQ2Pt, kZ) = static_cast<double>(getSigma1PtZ()) + static_cast<double>(rhs.getSigma1PtZ());
cov(kQ2Pt, kSnp) = static_cast<double>(getSigma1PtSnp()) + static_cast<double>(rhs.getSigma1PtSnp());
cov(kQ2Pt, kTgl) = static_cast<double>(getSigma1PtTgl()) + static_cast<double>(rhs.getSigma1PtTgl());
cov(kQ2Pt, kQ2Pt) = static_cast<double>(getSigma1Pt2()) + static_cast<double>(rhs.getSigma1Pt2());
}

//______________________________________________
template <typename value_T>
GPUd() bool TrackParametrizationWithError<value_T>::update(const TrackParametrizationWithError<value_T>& rhs, const MatrixDSym5& covInv)
Expand Down Expand Up @@ -867,7 +866,7 @@ GPUd() bool TrackParametrizationWithError<value_T>::update(const TrackParametriz
}

// updated covariance: Cov0 = Cov0 - K*Cov0
matK *= ROOT::Math::SMatrix<double, kNParams, kNParams, ROOT::Math::MatRepStd<double, kNParams>>(matC0);
matK *= o2::math_utils::SMatrix<double, kNParams, kNParams, o2::math_utils::MatRepStd<double, kNParams>>(matC0);
mC[kSigY2] -= matK(kY, kY);
mC[kSigZY] -= matK(kZ, kY);
mC[kSigZ2] -= matK(kZ, kZ);
Expand Down Expand Up @@ -901,8 +900,6 @@ GPUd() bool TrackParametrizationWithError<value_T>::update(const TrackParametriz
return update(rhs, covI);
}

#endif

//______________________________________________
template <typename value_T>
GPUd() bool TrackParametrizationWithError<value_T>::update(const value_t* p, const value_t* cov)
Expand Down Expand Up @@ -1245,6 +1242,7 @@ GPUd() void TrackParametrizationWithError<value_T>::printHexadecimal()

namespace o2::track
{

#if !defined(GPUCA_GPUCODE) || defined(GPUCA_GPUCODE_DEVICE) // FIXME: DR: WORKAROUND to avoid CUDA bug creating host symbols for device code.
template class TrackParametrizationWithError<float>;
#endif
Expand Down
26 changes: 18 additions & 8 deletions Detectors/ITSMFT/ITS/tracking/GPU/ITStrackingGPU/TimeFrameGPU.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -26,12 +26,9 @@

namespace o2
{
// namespace gpu
// {
// // class GPUChainITS;
// }
namespace its
{namespace gpu
{
namespace gpu
{

class DefaultGPUAllocator : public ExternalAllocator
Expand Down Expand Up @@ -59,13 +56,17 @@ class TimeFrameGPU : public TimeFrame
void loadClustersDevice();
void loadTrackletsDevice();
void loadCellsDevice();
void loadCellsLUT();
void loadTrackSeedsDevice();
void loadTrackSeedsChi2Device();
void loadRoadsDevice();
void loadTrackSeedsDevice(std::vector<CellSeed>&);
void createCellNeighboursDevice(const unsigned int& layer, std::vector<std::pair<int, int>>& neighbours);
void createNeighboursDevice(const unsigned int& layer, std::vector<std::pair<int, int>>& neighbours);
void createNeighboursLUTDevice(const int, const unsigned int);
void createTrackITSExtDevice(std::vector<CellSeed>&);
void downloadTrackITSExtDevice(std::vector<CellSeed>&);
void downloadCellsNeighbours(std::vector<std::vector<std::pair<int, int>>>&, const int);
void downloadNeighboursLUT(std::vector<int>&, const int);
void initDeviceChunks(const int, const int);
template <Task task>
size_t loadChunkData(const size_t, const size_t, const size_t);
Expand All @@ -92,17 +93,22 @@ class TimeFrameGPU : public TimeFrame
// Hybrid
Road<nLayers - 2>* getDeviceRoads() { return mRoadsDevice; }
TrackITSExt* getDeviceTrackITSExt() { return mTrackITSExtDevice; }
int* getDeviceNeighboursLUT(const int layer) { return mNeighboursLUTDevice[layer]; }
gpuPair<int, int>* getDeviceNeighbours(const int layer) { return mNeighboursDevice[layer]; }
TrackingFrameInfo* getDeviceTrackingFrameInfo(const int);
// TrackingFrameInfo** getDeviceArrayTrackingFrameInfo() { return mTrackingFrameInfoDeviceArray; }
const TrackingFrameInfo** getDeviceArrayTrackingFrameInfo() const { return mTrackingFrameInfoDeviceArray; }
Cluster** getDeviceArrayClusters() const { return mClustersDeviceArray; }
Cluster** getDeviceArrayUnsortedClusters() const { return mUnsortedClustersDeviceArray; }
Tracklet** getDeviceArrayTracklets() const { return mTrackletsDeviceArray; }
int** getDeviceArrayCellsLUT() const { return mCellsLUTDeviceArray; }
int** getDeviceArrayNeighboursCellLUT() const { return mNeighboursCellLUTDeviceArray; }
CellSeed** getDeviceArrayCells() const { return mCellsDeviceArray; }
CellSeed* getDeviceTrackSeeds() { return mTrackSeedsDevice; }
o2::track::TrackParCovF** getDeviceArrayTrackSeeds() { return mCellSeedsDeviceArray; }
float** getDeviceArrayTrackSeedsChi2() { return mCellSeedsChi2DeviceArray; }
int* getNeighboursIndexTablesDevice(const int layer) { return mNeighboursIndexTablesDevice[layer]; }

void setDevicePropagator(const o2::base::PropagatorImpl<float>*) override;

// Host-specific getters
Expand Down Expand Up @@ -131,9 +137,13 @@ class TimeFrameGPU : public TimeFrame
Cluster** mUnsortedClustersDeviceArray;
std::array<Tracklet*, nLayers - 1> mTrackletsDevice;
Tracklet** mTrackletsDeviceArray;
std::array<int*, nLayers - 2> mCellsLookupTablesDevice;
int** mCellsLookupTablesDeviceArray;
std::array<int*, nLayers - 2> mCellsLUTDevice;
std::array<int*, nLayers - 3> mNeighboursLUTDevice;
int** mCellsLUTDeviceArray;
int** mNeighboursCellDeviceArray;
int** mNeighboursCellLUTDeviceArray;
std::array<CellSeed*, nLayers - 2> mCellsDevice;
std::array<int*, nLayers - 2> mNeighboursIndexTablesDevice;
CellSeed* mTrackSeedsDevice;
CellSeed** mCellsDeviceArray;
std::array<o2::track::TrackParCovF*, nLayers - 2> mCellSeedsDevice;
Expand Down
27 changes: 27 additions & 0 deletions Detectors/ITSMFT/ITS/tracking/GPU/ITStrackingGPU/TrackingKernels.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -49,6 +49,33 @@ GPUg() void fitTrackSeedsKernel(
const o2::base::PropagatorF::MatCorrType matCorrType = o2::base::PropagatorF::MatCorrType::USEMatCorrLUT);
#endif
} // namespace gpu
void countCellNeighboursHandler(CellSeed** cellsLayersDevice,
int* neighboursLUTs,
int** cellsLUTs,
gpuPair<int, int>* cellNeighbours,
int* neighboursIndexTable,
const float maxChi2ClusterAttachment,
const float bz,
const int layerIndex,
const unsigned int nCells,
const unsigned int nCellsNext,
const int maxCellNeighbours,
const int nBlocks,
const int nThreads);

void computeCellNeighboursHandler(CellSeed** cellsLayersDevice,
int* neighboursLUTs,
int** cellsLUTs,
gpuPair<int, int>* cellNeighbours,
int* neighboursIndexTable,
const float maxChi2ClusterAttachment,
const float bz,
const int layerIndex,
const unsigned int nCells,
const unsigned int nCellsNext,
const int maxCellNeighbours,
const int nBlocks,
const int nThreads);

void trackSeedHandler(CellSeed* trackSeeds,
const TrackingFrameInfo** foundTrackingFrameInfo,
Expand Down
4 changes: 2 additions & 2 deletions Detectors/ITSMFT/ITS/tracking/GPU/cuda/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@
if(CUDA_ENABLED)
find_package(CUDAToolkit)
message(STATUS "Building ITS CUDA tracker")

add_compile_options(-O0 -g -fPIC)
o2_add_library(ITStrackingCUDA
SOURCES ClusterLinesGPU.cu
Context.cu
Expand All @@ -32,7 +32,7 @@ o2_add_library(ITStrackingCUDA
O2::SimulationDataFormat
O2::ReconstructionDataFormats
O2::GPUCommon
CUDA::nvToolsExt # TODO: change to CUDA::nvtx3 when CMake bump >= 3.25
CUDA::nvToolsExt
PRIVATE_LINK_LIBRARIES O2::GPUTrackingCUDAExternalProvider
TARGETVARNAME targetName)

Expand Down
Loading

0 comments on commit 5b6577c

Please sign in to comment.