Project
Loading...
Searching...
No Matches
GPUTPCDecompressionKernels.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
16#include "GPULogging.h"
17#include "GPUConstantMem.h"
19#include "GPUCommonAlgorithm.h"
20#include "TPCClusterDecompressionCore.inc"
21
22using namespace o2::gpu;
23using namespace o2::tpc;
24
25template <>
26GPUdii() void GPUTPCDecompressionKernels::Thread<GPUTPCDecompressionKernels::step0attached>(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUsharedref() GPUSharedMemory& smem, processorType& processors, int32_t trackStart, int32_t trackEnd)
27{
28 GPUTPCDecompression& GPUrestrict() decompressor = processors.tpcDecompressor;
29 CompressedClusters& GPUrestrict() cmprClusters = decompressor.mInputGPU;
30 const GPUParam& GPUrestrict() param = processors.param;
31
32 const uint32_t maxTime = (param.continuousMaxTimeBin + 1) * ClusterNative::scaleTimePacked - 1;
33
34 for (int32_t i = trackStart + get_global_id(0); i < trackEnd; i += get_global_size(0)) {
36 TPCClusterDecompressionCore::decompressTrack(cmprClusters, param, maxTime, i, offset, decompressor);
37 }
38}
39
40template <>
41GPUdii() void GPUTPCDecompressionKernels::Thread<GPUTPCDecompressionKernels::step1unattached>(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUsharedref() GPUSharedMemory& smem, processorType& processors, int32_t sectorStart, int32_t nSectors)
42{
43 GPUTPCDecompression& GPUrestrict() decompressor = processors.tpcDecompressor;
44 CompressedClusters& GPUrestrict() cmprClusters = decompressor.mInputGPU;
45 ClusterNative* GPUrestrict() clusterBuffer = decompressor.mNativeClustersBuffer;
46 const ClusterNativeAccess* outputAccess = decompressor.mClusterNativeAccess;
48 for (uint32_t i = get_global_id(0); i < GPUTPCGeometry::NROWS * nSectors; i += get_global_size(0)) {
49 uint32_t iRow = i % GPUTPCGeometry::NROWS;
50 uint32_t iSector = sectorStart + (i / GPUTPCGeometry::NROWS);
51 const uint32_t linearIndex = iSector * GPUTPCGeometry::NROWS + iRow;
52 uint32_t tmpBufferIndex = computeLinearTmpBufferIndex(iSector, iRow, decompressor.mMaxNativeClustersPerBuffer);
53 ClusterNative* buffer = clusterBuffer + outputAccess->clusterOffset[iSector][iRow];
54 if (decompressor.mNativeClustersIndex[linearIndex] != 0) {
55 decompressorMemcpyBasic(buffer, decompressor.mTmpNativeClusters + tmpBufferIndex, decompressor.mNativeClustersIndex[linearIndex]);
56 }
57 ClusterNative* clout = buffer + decompressor.mNativeClustersIndex[linearIndex];
58 uint32_t end = offsets[linearIndex] + ((linearIndex >= decompressor.mInputGPU.nSliceRows) ? 0 : decompressor.mInputGPU.nSliceRowClusters[linearIndex]);
59 TPCClusterDecompressionCore::decompressHits(cmprClusters, offsets[linearIndex], end, clout);
60 if (processors.param.rec.tpc.clustersEdgeFixDistance > 0.f) {
61 constexpr GPUTPCGeometry geo;
62 for (uint32_t k = 0; k < outputAccess->nClusters[iSector][iRow]; k++) {
63 auto& cluster = buffer[k];
64 if (cluster.getFlags() & ClusterNative::flagEdge) {
65 auto padF = cluster.getPad();
66 float distEdge = padF < geo.NPads(iRow) / 2 ? padF : geo.NPads(iRow) - 1 - padF;
67 if (distEdge > processors.param.rec.tpc.clustersEdgeFixDistance) {
68 cluster.setFlags(cluster.getFlags() ^ ClusterNative::flagEdge);
69 }
70 }
71 }
72 }
73 if (processors.param.rec.tpc.clustersShiftTimebins != 0.f) {
74 for (uint32_t k = 0; k < outputAccess->nClusters[iSector][iRow]; k++) {
75 auto& cl = buffer[k];
76 float t = cl.getTime() + processors.param.rec.tpc.clustersShiftTimebins;
77 if (t < 0) {
78 t = 0;
79 }
80 if (processors.param.continuousMaxTimeBin > 0 && t > processors.param.continuousMaxTimeBin) {
81 t = processors.param.continuousMaxTimeBin;
82 }
83 cl.setTime(t);
84 }
85 }
86 }
87}
88
89template <typename T>
90GPUdi() void GPUTPCDecompressionKernels::decompressorMemcpyBasic(T* GPUrestrict() dst, const T* GPUrestrict() src, uint32_t size)
91{
92 for (uint32_t i = 0; i < size; i++) {
93 dst[i] = src[i];
94 }
95}
96
98{
99 return param.tpcCutTimeBin > 0 ? cl.getTime() < param.tpcCutTimeBin : true;
100}
101
102template <>
103GPUdii() void GPUTPCDecompressionUtilKernels::Thread<GPUTPCDecompressionUtilKernels::countFilteredClusters>(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUsharedref() GPUSharedMemory& smem, processorType& processors)
104{
105 const GPUParam& GPUrestrict() param = processors.param;
106 GPUTPCDecompression& GPUrestrict() decompressor = processors.tpcDecompressor;
107 const ClusterNativeAccess* clusterAccess = decompressor.mClusterNativeAccess;
108 for (uint32_t i = get_global_id(0); i < GPUTPCGeometry::NSECTORS * GPUTPCGeometry::NROWS; i += get_global_size(0)) {
109 uint32_t sector = i / GPUTPCGeometry::NROWS;
110 uint32_t row = i % GPUTPCGeometry::NROWS;
111 for (uint32_t k = 0; k < clusterAccess->nClusters[sector][row]; k++) {
112 ClusterNative cl = clusterAccess->clusters[sector][row][k];
113 if (isClusterKept(cl, param)) {
114 decompressor.mNClusterPerSectorRow[i]++;
115 }
116 }
117 }
118}
119
120template <>
121GPUdii() void GPUTPCDecompressionUtilKernels::Thread<GPUTPCDecompressionUtilKernels::storeFilteredClusters>(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUsharedref() GPUSharedMemory& smem, processorType& processors)
122{
123 const GPUParam& GPUrestrict() param = processors.param;
124 GPUTPCDecompression& GPUrestrict() decompressor = processors.tpcDecompressor;
125 ClusterNative* GPUrestrict() clusterBuffer = decompressor.mNativeClustersBuffer;
126 const ClusterNativeAccess* clusterAccess = decompressor.mClusterNativeAccess;
127 const ClusterNativeAccess* outputAccess = processors.ioPtrs.clustersNative;
128 for (uint32_t i = get_global_id(0); i < GPUTPCGeometry::NSECTORS * GPUTPCGeometry::NROWS; i += get_global_size(0)) {
129 uint32_t sector = i / GPUTPCGeometry::NROWS;
130 uint32_t row = i % GPUTPCGeometry::NROWS;
131 uint32_t count = 0;
132 for (uint32_t k = 0; k < clusterAccess->nClusters[sector][row]; k++) {
133 const ClusterNative cl = clusterAccess->clusters[sector][row][k];
134 if (isClusterKept(cl, param)) {
135 clusterBuffer[outputAccess->clusterOffset[sector][row] + count] = cl;
136 count++;
137 }
138 }
139 }
140}
141
142template <>
143GPUdii() void GPUTPCDecompressionUtilKernels::Thread<GPUTPCDecompressionUtilKernels::sortPerSectorRow>(int32_t nBlocks, int32_t nThreads, int32_t iBlock, int32_t iThread, GPUsharedref() GPUSharedMemory& smem, processorType& processors)
144{
145 ClusterNative* GPUrestrict() clusterBuffer = processors.tpcDecompressor.mNativeClustersBuffer;
146 const ClusterNativeAccess* outputAccess = processors.ioPtrs.clustersNative;
147 for (uint32_t i = get_global_id(0); i < GPUTPCGeometry::NSECTORS * GPUTPCGeometry::NROWS; i += get_global_size(0)) {
148 uint32_t sector = i / GPUTPCGeometry::NROWS;
149 uint32_t row = i % GPUTPCGeometry::NROWS;
150 ClusterNative* buffer = clusterBuffer + outputAccess->clusterOffset[sector][row];
151 GPUCommonAlgorithm::sort(buffer, buffer + outputAccess->nClusters[sector][row]);
152 }
153}
int32_t i
#define GPUsharedref()
#define get_global_size(dim)
#define GPUrestrict()
#define get_global_id(dim)
uint32_t iSector
configStandalone Light OpenGL display display_light Apply smoothing to points Enable Z buffer Highlight links Highlight seeds Highlight sector tracks Highlight final tracks Show individual sector
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
GPUdii() void GPUTPCDecompressionKernels
o2::tpc::CompressedClusters mInputGPU
o2::tpc::ClusterNative * mTmpNativeClusters
o2::tpc::ClusterNative * mNativeClustersBuffer
o2::tpc::ClusterNativeAccess * mClusterNativeAccess
GLint GLsizei count
Definition glcorearb.h:399
GLsizeiptr size
Definition glcorearb.h:659
GLuint GLsizei const GLuint const GLintptr * offsets
Definition glcorearb.h:2595
GLuint GLuint end
Definition glcorearb.h:469
GLenum GLenum dst
Definition glcorearb.h:1767
GLintptr offset
Definition glcorearb.h:660
typedef void(APIENTRYP PFNGLCULLFACEPROC)(GLenum mode)
GLenum GLfloat param
Definition glcorearb.h:271
Global TPC definitions and constants.
GPUdi() T BetheBlochAleph(T bg
a couple of static helper functions to create timestamp values for CCDB queries or override obsolete ...
unsigned int nClusters[constants::MAXSECTOR][constants::MAXGLOBALPADROW]
const ClusterNative * clusters[constants::MAXSECTOR][constants::MAXGLOBALPADROW]
unsigned int clusterOffset[constants::MAXSECTOR][constants::MAXGLOBALPADROW]
std::vector< int > row