Project
Loading...
Searching...
No Matches
GPUChainTrackingTransformation.cxx
Go to the documentation of this file.
1// Copyright 2019-2020 CERN and copyright holders of ALICE O2.
2// See https://alice-o2.web.cern.ch/copyright for details of the copyright holders.
3// All rights not expressly granted are reserved.
4//
5// This software is distributed under the terms of the GNU General Public
6// License v3 (GPL Version 3), copied verbatim in the file "COPYING".
7//
8// In applying this license CERN does not waive the privileges and immunities
9// granted to it by virtue of its status as an Intergovernmental Organization
10// or submit itself to any jurisdiction.
11
14
15#include "GPUChainTracking.h"
16#include "GPULogging.h"
17#include "GPUO2DataTypes.h"
21#include "AliHLTTPCRawCluster.h"
22#include "GPUConstantMem.h"
23#include "GPUTPCClusterData.h"
24
28#include "utils/strtag.h"
29
30using namespace o2::gpu;
31using namespace o2::tpc;
32
33bool GPUChainTracking::NeedTPCClustersOnGPU()
34{
36}
37
39{
41 const auto& threadContext = GetThreadContext();
42
43 bool transferClusters = false;
44 if (mRec->IsGPU() && !(mRec->GetRecoStepsGPU() & gpudatatypes::RecoStep::TPCClusterFinding) && NeedTPCClustersOnGPU()) {
45 mInputsHost->mNClusterNative = mInputsShadow->mNClusterNative = mIOPtrs.clustersNative->nClustersTotal;
46 AllocateRegisteredMemory(mInputsHost->mResourceClusterNativeBuffer);
47 processorsShadow()->ioPtrs.clustersNative = mInputsShadow->mPclusterNativeAccess;
48 WriteToConstantMemory(RecoStep::TPCConversion, (char*)&processors()->ioPtrs - (char*)processors(), &processorsShadow()->ioPtrs, sizeof(processorsShadow()->ioPtrs), 0);
49 *mInputsHost->mPclusterNativeAccess = *mIOPtrs.clustersNative;
50 mInputsHost->mPclusterNativeAccess->clustersLinear = mInputsShadow->mPclusterNativeBuffer;
51 mInputsHost->mPclusterNativeAccess->setOffsetPtrs();
52 GPUMemCpy(RecoStep::TPCConversion, mInputsShadow->mPclusterNativeBuffer, mIOPtrs.clustersNative->clustersLinear, sizeof(mIOPtrs.clustersNative->clustersLinear[0]) * mIOPtrs.clustersNative->nClustersTotal, 0, true);
53 TransferMemoryResourceLinkToGPU(RecoStep::TPCConversion, mInputsHost->mResourceClusterNativeAccess, 0);
54 transferClusters = true;
55 }
56 if (GetProcessingSettings().debugLevel >= 3) {
57 GPUInfo("Early transform inactive, skipping TPC Early transformation kernel, transformed on the fly during sector data creation / refit");
58 }
59 if (transferClusters) {
60 SynchronizeStream(0); // TODO: Synchronize implicitly with next step
61 }
62 return 0;
63}
64
66{
68 for (uint32_t i = 0; i < NSECTORS; i++) {
69 mIOPtrs.rawClusters[i] = nullptr;
71 mIOMem.rawClusters[i].reset(nullptr);
72 mIOPtrs.clusterData[i] = nullptr;
74 mIOMem.clusterData[i].reset(nullptr);
75 }
77 if (GetProcessingSettings().registerStandaloneInputMemory) {
78 if (mRec->registerMemoryForGPU(mIOMem.clustersNative.get(), mIOMem.clusterNativeAccess->nClustersTotal * sizeof(*mIOMem.clusterNativeAccess->clustersLinear))) {
79 throw std::runtime_error("Error registering memory for GPU");
80 }
81 }
82}
83
85{
92 if (GetProcessingSettings().registerStandaloneInputMemory) {
93 for (uint32_t i = 0; i < NSECTORS; i++) {
94 for (uint32_t j = 0; j < GPUTrackingInOutZS::NENDPOINTS; j++) {
95 for (uint32_t k = 0; k < mIOPtrs.tpcZS->sector[i].count[j]; k++) {
96 if (mRec->registerMemoryForGPU(mIOPtrs.tpcZS->sector[i].zsPtr[j][k], mIOPtrs.tpcZS->sector[i].nZSPtr[j][k] * TPCZSHDR::TPC_ZS_PAGE_SIZE)) {
97 throw std::runtime_error("Error registering memory for GPU");
98 }
99 }
100 }
101 }
102 }
103}
104
109
111{
112 if (GetRecoStepsGPU() & RecoStep::TPCClusterFinding) {
113 throw std::runtime_error("Cannot forward TPC digits with Clusterizer on GPU");
114 }
115 std::vector<ClusterNative> tmp[NSECTORS][GPUTPCGeometry::NROWS];
116 uint32_t nTotal = 0;
117 const float zsThreshold = param().rec.tpc.zsThreshold;
118 for (int32_t i = 0; i < NSECTORS; i++) {
119 for (uint32_t j = 0; j < mIOPtrs.tpcPackedDigits->nTPCDigits[i]; j++) {
120 const auto& d = mIOPtrs.tpcPackedDigits->tpcDigits[i][j];
121 if (d.getChargeFloat() >= zsThreshold) {
123 c.setTimeFlags(d.getTimeStamp(), 0);
124 c.setPad(d.getPad());
125 c.setSigmaTime(1);
126 c.setSigmaPad(1);
127 c.qTotPacked = c.qMax = d.getChargeFloat();
128 tmp[i][d.getRow()].emplace_back(c);
129 nTotal++;
130 }
131 }
132 }
133 mIOMem.clustersNative.reset(new ClusterNative[nTotal]);
134 nTotal = 0;
135 mClusterNativeAccess->clustersLinear = mIOMem.clustersNative.get();
136 for (uint32_t i = 0; i < NSECTORS; i++) {
137 for (uint32_t j = 0; j < GPUTPCGeometry::NROWS; j++) {
138 mClusterNativeAccess->nClusters[i][j] = tmp[i][j].size();
139 memcpy(&mIOMem.clustersNative[nTotal], tmp[i][j].data(), tmp[i][j].size() * sizeof(*mClusterNativeAccess->clustersLinear));
140 nTotal += tmp[i][j].size();
141 }
142 }
143 mClusterNativeAccess->setOffsetPtrs();
144 mIOPtrs.tpcPackedDigits = nullptr;
146 GPUInfo("Forwarded %u TPC clusters", nTotal);
147 mRec->MemoryScalers()->nTPCHits = nTotal;
148 return 0;
149}
Class of a TPC cluster in TPC-native coordinates (row, time)
unsigned int uint32_t
int32_t i
Timeframe settings Average distance in cm of events merged into Randomize distance around average distance of merged events Simulate correct bunch interactions instead of placing only the average number of events A value[n] sets TFlen for[n] collisions in Number of bunches per trainsort Number of bunch trains Instantaneous interaction rate Do not simulate border Total number of collisions to be placed in the interior of all time Do not select random but walk over array of events in stride steps h Event generator settings h configStandalone Use CPU for GPU Force usage of the specified GPU device no CPU fallback Continue processing after an error First event to Number of events to Number of iterations to Number of iterations to Number of initial iterations excluded from const char Directory with events to Events directory is and not inside events d
configStandalone Reconstruction rec_tpc multiplier for the combined cluster track error during track following area in cm for the search of for z only used if decrease removal by when above this local occupancy rowx Shift of TPC revert cluster flag edge bit distance to edge exceeds this ZOffset for secondary multiple of radius max P of electron K BB bands crossing max P of electron p BB bands crossing max P of electron d BB bands crossing max P of electron t BB bands crossing min accepted dEdxMaxTPC of the track Additive extra cluster error for Y2 if edge flag set Additive extra cluster error for Y2 if single set Additive extra cluster error for Y2 if or single set Additive extra cluster error for Z2 if or single set Extra errors to add to track when crossing depending on addErrorsCECrossing Inner portion of y range in sector that is not used in searching for extrapolated track candidates factor2YS for track merging factor2K for track merging Margin in cm of Y position when rejecting edge clusters based on uncorrected track Y Max fraction of shared clusters for tracklet Except pad from cluster finding if total number of charges in a fragment is above this Except pad from cluster finding if number of consecutive charges in a fragment is above this Threshold where a timebin is considered disabling the noisy pad check for that pad Treat signal above this minimum as saturated Smoothing factor for the exponential Highly Ionising Particle tail filter Number of timebins to use for the averaging Add additional custom track errors when crossing relative as fraction of Do N fit passes in final fit of< 0:reject at max -n hits") AddOptionRTC(rejectIFCLowRadiusCluster, uint8_t, 1, "", 0, "Reject clusters that get the IFC mask error during refit") AddOptionRTC(dEdxTruncLow, uint8_t, 2, "", 0, "Low truncation threshold, fraction of 128") AddOptionRTC(dEdxTruncHigh, uint8_t, 77, "", 0, "High truncation threshold, fraction of 128") AddOptionRTC(extrapolationTracking, int8_t, 1, "", 0, "Enable Extrapolation Tracking(prolong tracks to adjacent sectors to find short segments)") AddOptionRTC(disableRefitAttachment, uint8_t, 0, "", 0, "Bitmask to disable certain attachment steps during refit(1:attachment, 2:propagation, 4:loop following, 8:mirroring)") AddOptionRTC(rejectionStrategy, uint8_t, o2::gpu::GPUSettings::RejectionStrategyA, "", 0, "Enable rejection of TPC clusters for compression(0=no, 1=strategy A, 2=strategy B)") AddOptionRTC(mergeLoopersAfterburner, uint8_t, 1, "", 0, "Run afterburner for additional looper merging") AddOptionRTC(compressionTypeMask, uint8_t, o2::gpu::GPUSettings::CompressionFull, "", 0, "TPC Compression mode bits(1=truncate charge/width LSB, 2=differences, 4=track-model)") AddOptionRTC(compressionSortOrder, uint8_t, o2::gpu::GPUSettings::SortTime, "", 0, "Sort order of TPC compression(0=time, 1=pad, 2=Z-time-pad, 3=Z-pad-time, 4=no sorting(use incoming order))") AddOptionRTC(sigBitsCharge, uint8_t, 4, "", 0, "Number of significant bits for TPC cluster charge in compression mode 1") AddOptionRTC(sigBitsWidth, uint8_t, 3, "", 0, "Number of significant bits for TPC cluster width in compression mode 1") AddOptionRTC(dropLoopers, uint8_t, 0, "", 0, "Drop looping tracks starting from second loop") AddOptionRTC(mergerInterpolateErrors, uint8_t, 1, "", 0, "Use interpolation instead of extrapolation for chi2 based cluster rejection") AddOptionRTC(mergerInterpolateRejectAlsoOnCurrentPosition, uint8_t, 1, "", 0, "When using mergerInterpolateErrors, reject based on chi2 twice computed with interpolated and current track position starting from NDF > mergerNonInterpolateRejectMinNDF Merge tracks accross the central electrode Enable PID response Add sub threshold clusters in TPC dEdx computation OR mask of TPC flags that will reject the cluster in alternative dEdx Reject edge clusters based on uncorrected track Y during track fit Increase decrease removal by when on this number of pads from the edge h configStandalone Reconstruction rec_trd Max chi2 for TRD tracklets to be matched to a track Chi2 cut for strict matching mode Number of sigmas for ITS TPC track time error estimate Addition to search road around track prolongation along Y in cm Optimal tracklet rphi resolution in maximum chi2 per attached tracklet for TRD tracks take into account the full likelihood of the slope including exponential tails Tracks with less attached tracklets are discarded after the tracking Post trigger Pile up integration time in BCs fits all iterations but the last one in projections Use material lookup table for TPC refit Apply workarounds to avoid cyclic merge should not be needed tpc dyn configStandalone Processing proc_rtc Replace constant variables static by constexpr expressions Compile RTC in deterministic with NO_FAST_MATH flags and GPUCA_DETERMINISTIC_MODE define Use RTC to optimize GPU code h configStandalone Processing proc_rtctech Use a file lock to serialize access to the cache folder int32_t
uint32_t j
Definition RawData.h:0
uint32_t c
Definition RawData.h:2
uint32_t version
Definition RawData.h:8
Definitions of TPC Zero Suppression Data Headers.
std::unique_ptr< o2::tpc::ClusterNativeAccess > mClusterNativeAccess
std::unique_ptr< GPUTrackingInputProvider > mInputsHost
GPUTrackingInOutPointers & mIOPtrs
struct o2::gpu::GPUChainTracking::InOutMemory mIOMem
std::unique_ptr< GPUTrackingInputProvider > mInputsShadow
void TransferMemoryResourceLinkToGPU(RecoStep step, int16_t res, int32_t stream=-1, deviceEvent *ev=nullptr, deviceEvent *evList=nullptr, int32_t nEvents=1)
Definition GPUChain.h:124
void GPUMemCpy(RecoStep step, void *dst, const void *src, size_t size, int32_t stream, int32_t toGPU, deviceEvent *ev=nullptr, deviceEvent *evList=nullptr, int32_t nEvents=1)
Definition GPUChain.h:129
GPUReconstruction::RecoStepField GetRecoStepsGPU() const
Definition GPUChain.h:72
void WriteToConstantMemory(RecoStep step, size_t offset, const void *src, size_t size, int32_t stream=-1, deviceEvent *ev=nullptr)
Definition GPUChain.h:128
size_t AllocateRegisteredMemory(GPUProcessor *proc)
Definition GPUChain.h:228
virtual std::unique_ptr< GPUReconstructionProcessing::threadContext > GetThreadContext()
Definition GPUChain.h:109
GPUConstantMem * processors()
Definition GPUChain.h:84
GPUParam & param()
Definition GPUChain.h:87
const GPUSettingsProcessing & GetProcessingSettings() const
Definition GPUChain.h:76
void SynchronizeStream(int32_t stream)
Definition GPUChain.h:89
GPUReconstructionCPU * mRec
Definition GPUChain.h:79
GPUConstantMem * processorsShadow()
Definition GPUChain.h:85
static constexpr int32_t NSECTORS
Definition GPUChain.h:58
GPUReconstruction * rec()
Definition GPUChain.h:66
static void RunZSEncoder(const S &in, std::unique_ptr< uint64_t[]> *outBuffer, uint32_t *outSizes, o2::raw::RawFileWriter *raw, const o2::InteractionRecord *ir, const GPUParam &param, int32_t version, bool verify, float threshold=0.f, bool padding=false, std::function< void(std::vector< o2::tpc::Digit > &)> digitsFilter=nullptr)
static void RunZSFilter(std::unique_ptr< o2::tpc::Digit[]> *buffers, const o2::tpc::Digit *const *ptrs, size_t *nsb, const size_t *ns, const GPUParam &param, bool zs12bit, float threshold)
static void RunZSEncoderCreateMeta(const uint64_t *buffer, const uint32_t *sizes, void **ptrs, GPUTrackingInOutZS *out)
static void ConvertRun2RawToNative(o2::tpc::ClusterNativeAccess &native, std::unique_ptr< o2::tpc::ClusterNative[]> &nativeBuffer, const AliHLTTPCRawCluster **rawClusters, uint32_t *nRawClusters)
RecoStepField GetRecoStepsGPU() const
void PushNonPersistentMemory(uint64_t tag)
GPUMemorySizeScalers * MemoryScalers()
int32_t registerMemoryForGPU(const void *ptr, size_t size)
GLenum GLfloat param
Definition glcorearb.h:271
Global TPC definitions and constants.
constexpr T qStr2Tag(const char(&str)[N])
Definition strtag.h:24
std::unique_ptr< o2::tpc::Digit[]> tpcDigits[NSECTORS]
std::unique_ptr< GPUTrackingInOutDigits > digitMap
std::unique_ptr< GPUTPCClusterData[]> clusterData[NSECTORS]
std::unique_ptr< o2::tpc::ClusterNative[]> clustersNative
std::unique_ptr< uint64_t[]> tpcZSpages
std::unique_ptr< GPUTrackingInOutZS > tpcZSmeta
std::unique_ptr< o2::tpc::ClusterNativeAccess > clusterNativeAccess
std::unique_ptr< AliHLTTPCRawCluster[]> rawClusters[NSECTORS]
std::unique_ptr< GPUTrackingInOutZS::GPUTrackingInOutZSMeta > tpcZSmeta2
GPUTrackingInOutPointers ioPtrs
const o2::tpc::Digit * tpcDigits[NSECTORS]
const o2::tpc::ClusterNativeAccess * clustersNative
const GPUTrackingInOutZS * tpcZS
const AliHLTTPCRawCluster * rawClusters[NSECTORS]
const GPUTPCClusterData * clusterData[NSECTORS]
const GPUTrackingInOutDigits * tpcPackedDigits
GPUTrackingInOutZSSector sector[NSECTORS]
const ClusterNative * clustersLinear
o2::InteractionRecord ir(0, 0)