#include <cstdint>
#include <cuda_runtime.h>
#include "CUDADataFormats/TrackingRecHit/interface/TrackingRecHitsUtilities.h"
#include "HeterogeneousCore/CUDAUtilities/interface/cudaCheck.h"
#include "HeterogeneousCore/CUDAUtilities/interface/cuda_assert.h"
#include "RecoLocalTracker/SiPixelRecHits/interface/pixelCPEforGPU.h"
#include "RecoTracker/PixelTrackFitting/interface/BrokenLine.h"
#include "HelixFitOnGPU.h"

Typedefs
template<typename TrackerTraits >
using	OutputSoAView = TrackSoAView< TrackerTraits >

template<typename TrackerTraits >
using	TupleMultiplicity = caStructures::TupleMultiplicityT< TrackerTraits >

template<typename TrackerTraits >
using	Tuples = typename TrackSoA< TrackerTraits >::HitContainer

Functions
template<int N, typename TrackerTraits >
	__attribute__ ((always_inline)) void kernel_BLFastFit(Tuples< TrackerTraits > const *__restrict__ foundNtuplets

	assert (hitsInFit<=nHitsL)

	assert (nHitsL<=nHitsH)

	assert (phits)

	assert (pfast_fit)

	assert (foundNtuplets)

	assert (tupleMultiplicity)

	assert (totTK<=int(tupleMultiplicity->size()))

	assert (totTK >=0)

	assert (results_view.eta())

	for (int local_idx=local_start, nt=riemannFit::maxNumberOfConcurrentFits;local_idx< nt;local_idx+=gridDim.x *blockDim.x)

Variables
double	bField

TupleMultiplicity< TrackerTraits > const *__restrict__ TrackingRecHitSoAConstView< TrackerTraits >	hh

constexpr auto	invalidTkId = std::numeric_limits<typename TrackerTraits::tindex_type>::max()

auto	local_start = blockIdx.x * blockDim.x + threadIdx.x

TupleMultiplicity< TrackerTraits > const __restrict__ TrackingRecHitSoAConstView< TrackerTraits > TrackerTraits::tindex_type __restrict__ double __restrict__ float __restrict__ double *__restrict__ uint32_t uint32_t	nHitsH

TupleMultiplicity< TrackerTraits > const __restrict__ TrackingRecHitSoAConstView< TrackerTraits > TrackerTraits::tindex_type __restrict__ double __restrict__ float __restrict__ double *__restrict__ uint32_t	nHitsL

TupleMultiplicity< TrackerTraits > const __restrict__ TrackingRecHitSoAConstView< TrackerTraits > TrackerTraits::tindex_type __restrict__ double __restrict__ float __restrict__ double *__restrict__ uint32_t uint32_t int32_t	offset

TupleMultiplicity< TrackerTraits > const __restrict__ TrackingRecHitSoAConstView< TrackerTraits > TrackerTraits::tindex_type __restrict__ double __restrict__ float __restrict__ double *__restrict__	pfast_fit

TupleMultiplicity< TrackerTraits > const __restrict__ TrackingRecHitSoAConstView< TrackerTraits > TrackerTraits::tindex_type __restrict__ double *__restrict__	phits

TupleMultiplicity< TrackerTraits > const __restrict__ TrackingRecHitSoAConstView< TrackerTraits > TrackerTraits::tindex_type __restrict__ double __restrict__ float __restrict__	phits_ge

TupleMultiplicity< TrackerTraits > const __restrict__ TrackingRecHitSoAConstView< TrackerTraits > TrackerTraits::tindex_type __restrict__	ptkids

double OutputSoAView< TrackerTraits >	results_view

int	totTK = tupleMultiplicity->end(nHitsH) - tupleMultiplicity->begin(nHitsL)

TupleMultiplicity< TrackerTraits > const *__restrict__	tupleMultiplicity

Typedef Documentation

◆ OutputSoAView

template<typename TrackerTraits >

using OutputSoAView = TrackSoAView<TrackerTraits>

Definition at line 22 of file BrokenLineFitOnGPU.h.

◆ TupleMultiplicity

template<typename TrackerTraits >

using TupleMultiplicity = caStructures::TupleMultiplicityT<TrackerTraits>

Definition at line 24 of file BrokenLineFitOnGPU.h.

◆ Tuples

template<typename TrackerTraits >

using Tuples = typename TrackSoA<TrackerTraits>::HitContainer

Definition at line 20 of file BrokenLineFitOnGPU.h.

Function Documentation

◆ attribute()

template<int N, typename TrackerTraits >

__attribute__ ( (always_inline) ) const

inline

◆ assert() [1/9]

assert ( hitsInFit<= nHitsL )

Referenced by for().

◆ assert() [2/9]

assert ( nHitsL<= nHitsH )

◆ assert() [3/9]

assert ( phits )

◆ assert() [4/9]

assert ( pfast_fit )

◆ assert() [5/9]

assert ( foundNtuplets )

◆ assert() [6/9]

assert ( tupleMultiplicity )

◆ assert() [7/9]

assert ( totTK<= inttupleMultiplicity->size() )

◆ assert() [8/9]

assert ( totTK >= 0 )

◆ assert() [9/9]

assert ( results_view. eta() )

◆ for()

for ( int local_idx = local_start )

Definition at line 62 of file BrokenLineFitOnGPU.h.

References cms::cudacompat::__syncthreads(), funct::abs(), assert(), cms::cudacompat::atomicAdd(), newFWLiteAna::bin, fileCollector::done, Calorimetry_cff::dp, GCP_Ntuples_cfg::dump, PVValHelper::dx, PVValHelper::dy, PVValHelper::dz, f, ALPAKA_ACCELERATOR_NAMESPACE::brokenline::fast_fit, brokenline::fastFit(), caHitNtupletGeneratorKernels::foundNtuplets, hh, hfClusterShapes_cfi::hits, mps_fire::i, createfilelist::int, invalidTkId, dqmiolumiharvest::j, CPEFastParametrisation::kGenErrorQBins, CPEFastParametrisation::kNumErrorBins, SiStripPI::max, dqmiodumpmetadata::n, nHits, nHitsH, nHitsL, offset, pfast_fit, phits, phits_ge, ptkids, mps_update::status, totTK, and tupleMultiplicity.

                                             {
     int tuple_idx = local_idx + offset;
     if (tuple_idx >= totTK) {
       ptkids[local_idx] = invalidTkId;
       break;
     }
     // get it from the ntuple container (one to one to helix)
     auto tkid = *(tupleMultiplicity->begin(nHitsL) + tuple_idx);
     assert(int(tkid) < foundNtuplets->nOnes());
 
     ptkids[local_idx] = tkid;
 
     auto nHits = foundNtuplets->size(tkid);
 
     assert(nHits >= nHitsL);
     assert(nHits <= nHitsH);
 
     riemannFit::Map3xNd<N> hits(phits + local_idx);
     riemannFit::Map4d fast_fit(pfast_fit + local_idx);
     riemannFit::Map6xNf<N> hits_ge(phits_ge + local_idx);
 
 #ifdef BL_DUMP_HITS
     __shared__ int done;
     done = 0;
     __syncthreads();
     bool dump = (foundNtuplets->size(tkid) == 5 && 0 == atomicAdd(&done, 1));
 #endif
 
     // Prepare data structure
     auto const *hitId = foundNtuplets->begin(tkid);
 
     // #define YERR_FROM_DC
 #ifdef YERR_FROM_DC
     // try to compute more precise error in y
     auto dx = hh[hitId[hitsInFit - 1]].xGlobal() - hh[hitId[0]].xGlobal();
     auto dy = hh[hitId[hitsInFit - 1]].yGlobal() - hh[hitId[0]].yGlobal();
     auto dz = hh[hitId[hitsInFit - 1]].zGlobal() - hh[hitId[0]].zGlobal();
     float ux, uy, uz;
 #endif
 
     float incr = std::max(1.f, float(nHits) / float(hitsInFit));
     float n = 0;
     for (uint32_t i = 0; i < hitsInFit; ++i) {
       int j = int(n + 0.5f);  // round
       if (hitsInFit - 1 == i)
         j = nHits - 1;  // force last hit to ensure max lever arm.
       assert(j < int(nHits));
       n += incr;
       auto hit = hitId[j];
       float ge[6];
 
 #ifdef YERR_FROM_DC
       auto const &dp = hh.cpeParams().detParams(hh.detectorIndex(hit));
       auto status = hh[hit].chargeAndStatus().status;
       int qbin = CPEFastParametrisation::kGenErrorQBins - 1 - status.qBin;
       assert(qbin >= 0 && qbin < 5);
       bool nok = (status.isBigY | status.isOneY);
       // compute cotanbeta and use it to recompute error
       dp.frame.rotation().multiply(dx, dy, dz, ux, uy, uz);
       auto cb = std::abs(uy / uz);
       int bin =
           int(cb * (float(phase1PixelTopology::pixelThickess) / float(phase1PixelTopology::pixelPitchY)) * 8.f) - 4;
       int low_value = 0;
       int high_value = CPEFastParametrisation::kNumErrorBins - 1;
       // return estimated bin value truncated to [0, 15]
       bin = std::clamp(bin, low_value, high_value);
       float yerr = dp.sigmay[bin] * 1.e-4f;  // toCM
       yerr *= dp.yfact[qbin];                // inflate
       yerr *= yerr;
       yerr += dp.apeYY;
       yerr = nok ? hh[hit].yerrLocal() : yerr;
       dp.frame.toGlobal(hh[hit].xerrLocal(), 0, yerr, ge);
 #else
       hh.cpeParams().detParams(hh[hit].detectorIndex()).frame.toGlobal(hh[hit].xerrLocal(), 0, hh[hit].yerrLocal(), ge);
 #endif
 
 #ifdef BL_DUMP_HITS
       bool dump = foundNtuplets->size(tkid) == 5;
       if (dump) {
         printf("Track id %d %d Hit %d on %d\nGlobal: hits.col(%d) << %f,%f,%f\n",
                local_idx,
                tkid,
                hit,
                hh[hit].detectorIndex(),
                i,
                hh[hit].xGlobal(),
                hh[hit].yGlobal(),
                hh[hit].zGlobal());
         printf("Error: hits_ge.col(%d) << %e,%e,%e,%e,%e,%e\n", i, ge[0], ge[1], ge[2], ge[3], ge[4], ge[5]);
       }
 #endif
 
       hits.col(i) << hh[hit].xGlobal(), hh[hit].yGlobal(), hh[hit].zGlobal();
       hits_ge.col(i) << ge[0], ge[1], ge[2], ge[3], ge[4], ge[5];
     }
     brokenline::fastFit(hits, fast_fit);
 
     // no NaN here....
     assert(fast_fit(0) == fast_fit(0));
     assert(fast_fit(1) == fast_fit(1));
     assert(fast_fit(2) == fast_fit(2));
     assert(fast_fit(3) == fast_fit(3));
   }

Variable Documentation

◆ bField

uint32_t double bField

Definition at line 170 of file BrokenLineFitOnGPU.h.

◆ hh

TupleMultiplicity<TrackerTraits> const* __restrict__ TrackingRecHitSoAConstView<TrackerTraits> hh

Definition at line 30 of file BrokenLineFitOnGPU.h.

Referenced by for().

◆ invalidTkId

constexpr auto invalidTkId = std::numeric_limits<typename TrackerTraits::tindex_type>::max()

Definition at line 40 of file BrokenLineFitOnGPU.h.

Referenced by for(), ALPAKA_ACCELERATOR_NAMESPACE::Kernel_BLFastFit< N, TrackerTraits >::operator()(), and ALPAKA_ACCELERATOR_NAMESPACE::Kernel_BLFit< N, TrackerTraits >::operator()().

◆ local_start

auto local_start = blockIdx.x * blockDim.x + threadIdx.x

Definition at line 50 of file BrokenLineFitOnGPU.h.

◆ nHitsH

TupleMultiplicity<TrackerTraits> const* __restrict__ TrackingRecHitSoAConstView<TrackerTraits> TrackerTraits::tindex_type* __restrict__ double* __restrict__ float* __restrict__ double* __restrict__ uint32_t uint32_t nHitsH

Definition at line 30 of file BrokenLineFitOnGPU.h.

Referenced by for(), and ALPAKA_ACCELERATOR_NAMESPACE::Kernel_BLFastFit< N, TrackerTraits >::operator()().

◆ nHitsL

TupleMultiplicity<TrackerTraits> const* __restrict__ TrackingRecHitSoAConstView<TrackerTraits> TrackerTraits::tindex_type* __restrict__ double* __restrict__ float* __restrict__ double* __restrict__ uint32_t nHitsL

Definition at line 30 of file BrokenLineFitOnGPU.h.

Referenced by for(), and ALPAKA_ACCELERATOR_NAMESPACE::Kernel_BLFastFit< N, TrackerTraits >::operator()().

◆ offset

uint32_t double OutputSoAView< TrackerTraits > double *__restrict__ float *__restrict__ double *__restrict__ riemannFit::CircleFit *__restrict__ uint32_t offset

Initial value:

{

constexpr uint32_t hitsInFit = N

ALPAKA_ACCELERATOR_NAMESPACE::brokenline::constexpr

if constexpr(n > 3)

Definition: BrokenLine.h:165

N

#define N

Definition: blowfish.cc:9

Definition at line 38 of file BrokenLineFitOnGPU.h.

Referenced by for().

◆ pfast_fit

double OutputSoAView< TrackerTraits > TrackerTraits::tindex_type const *__restrict__ double *__restrict__ float *__restrict__ double *__restrict__ pfast_fit

Initial value:

{

assert(results_view.pt())

results_view

double OutputSoAView< TrackerTraits > results_view

Definition: BrokenLineFitOnGPU.h:170

assert

assert(hitsInFit<=nHitsL)

Definition at line 30 of file BrokenLineFitOnGPU.h.

Referenced by for(), ALPAKA_ACCELERATOR_NAMESPACE::Kernel_BLFastFit< N, TrackerTraits >::operator()(), ALPAKA_ACCELERATOR_NAMESPACE::Kernel_FastFit< N, TrackerTraits >::operator()(), and ALPAKA_ACCELERATOR_NAMESPACE::Kernel_BLFit< N, TrackerTraits >::operator()().