Project
Loading...
Searching...
No Matches
GPUReconstructionCPU.h
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#ifndef GPURECONSTRUCTIONICPU_H
16#define GPURECONSTRUCTIONICPU_H
17
19#include <fstream>
20#include <stdexcept>
21#include <string>
22#include <vector>
23
24namespace Ort
25{
26struct SessionOptions;
27}
28
29namespace o2::gpu
30{
31
32class GPUReconstructionCPU : public GPUReconstructionProcessing::KernelInterface<GPUReconstructionCPU, GPUReconstructionProcessing>
33{
35 friend class GPUChain;
36
37 public:
38 ~GPUReconstructionCPU() override;
39 static constexpr krnlRunRange krnlRunRangeNone{0};
40 static constexpr krnlEvent krnlEventNone = krnlEvent{nullptr, nullptr, 0};
41
42 template <class S, int32_t I = 0>
44 template <class T, int32_t I = 0, typename... Args>
45 void runKernelBackend(const krnlSetupTime& _xyz, const Args&... args);
46
47 virtual int32_t GPUDebug(const char* state = "UNKNOWN", int32_t stream = -1, bool force = false);
50
51 int32_t RunChains() override;
52
53 void UpdateParamOccupancyMap(const uint32_t* mapHost, const uint32_t* mapGPU, uint32_t occupancyTotal, uint32_t mapSize, int32_t stream = -1, deviceEvent* ev = nullptr);
54
55 protected:
61
63
64 int32_t registerMemoryForGPU_internal(const void* ptr, size_t size) override { return 0; }
65 int32_t unregisterMemoryForGPU_internal(const void* ptr) override { return 0; }
66
68 virtual void SynchronizeEvents(deviceEvent* evList, int32_t nEvents = 1) {}
70 virtual bool IsEventDone(deviceEvent* evList, int32_t nEvents = 1) { return true; }
71 virtual void RecordMarker(deviceEvent* ev, int32_t stream) {}
72 virtual void SynchronizeGPU() {}
73 virtual void ReleaseEvent(deviceEvent ev) {}
74
75 size_t TransferMemoryResourceToGPU(GPUMemoryResource* res, int32_t stream = -1, deviceEvent* ev = nullptr, deviceEvent* evList = nullptr, int32_t nEvents = 1) { return TransferMemoryInternal(res, stream, ev, evList, nEvents, true, res->Ptr(), res->PtrDevice()); }
76 size_t TransferMemoryResourceToHost(GPUMemoryResource* res, int32_t stream = -1, deviceEvent* ev = nullptr, deviceEvent* evList = nullptr, int32_t nEvents = 1) { return TransferMemoryInternal(res, stream, ev, evList, nEvents, false, res->PtrDevice(), res->Ptr()); }
77 size_t TransferMemoryResourcesToGPU(GPUProcessor* proc, int32_t stream = -1, bool all = false) { return TransferMemoryResourcesHelper(proc, stream, all, true); }
78 size_t TransferMemoryResourcesToHost(GPUProcessor* proc, int32_t stream = -1, bool all = false) { return TransferMemoryResourcesHelper(proc, stream, all, false); }
79 size_t TransferMemoryResourceLinkToGPU(int16_t res, int32_t stream = -1, deviceEvent* ev = nullptr, deviceEvent* evList = nullptr, int32_t nEvents = 1) { return TransferMemoryResourceToGPU(&mMemoryResources[res], stream, ev, evList, nEvents); }
80 size_t TransferMemoryResourceLinkToHost(int16_t res, int32_t stream = -1, deviceEvent* ev = nullptr, deviceEvent* evList = nullptr, int32_t nEvents = 1) { return TransferMemoryResourceToHost(&mMemoryResources[res], stream, ev, evList, nEvents); }
81 virtual size_t GPUMemCpy(void* dst, const void* src, size_t size, int32_t stream, int32_t toGPU, deviceEvent* ev = nullptr, deviceEvent* evList = nullptr, int32_t nEvents = 1);
82 virtual size_t GPUMemCpyAlways(bool onGpu, void* dst, const void* src, size_t size, int32_t stream, int32_t toGPU, deviceEvent* ev = nullptr, deviceEvent* evList = nullptr, int32_t nEvents = 1);
83 size_t WriteToConstantMemory(size_t offset, const void* src, size_t size, int32_t stream = -1, deviceEvent* ev = nullptr) override;
84 virtual size_t TransferMemoryInternal(GPUMemoryResource* res, int32_t stream, deviceEvent* ev, deviceEvent* evList, int32_t nEvents, bool toGPU, const void* src, void* dst);
85
86 // ONNX runtime
87 virtual void SetONNXGPUStream(Ort::SessionOptions&, int32_t, int32_t*) {}
88
89 int32_t InitDevice() override;
90 int32_t ExitDevice() override;
92
93 // Pointers to tracker classes
94 GPUProcessorProcessors mProcShadow; // Host copy of tracker objects that will be used on the GPU
96
100
101 private:
102 size_t TransferMemoryResourcesHelper(GPUProcessor* proc, int32_t stream, bool all, bool toGPU);
103 template <class S, int32_t I = 0, typename... Args>
104 void runKernelInterface(krnlSetup&& setup, Args const&... args);
105
106 struct debugWriter {
107 debugWriter(std::string filenameCSV, bool markdown, uint32_t statNEvents);
108 void header();
109 void row(char type, uint32_t count, std::string name, double gpu_time, double cpu_time, double total_time, std::size_t memSize, std::string nEventReport = "");
110
111 private:
112 std::ofstream streamCSV;
113 bool mMarkdown;
114 uint32_t mStatNEvents;
115 };
116};
117
118} // namespace o2::gpu
119
120#endif
std::vector< std::string > header
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
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 proc
SurfaceTrackState state
uint32_t res
Definition RawData.h:0
TBranch * ptr
GPUReconstructionCPU(const GPUSettingsDeviceBackend &cfg)
virtual void SetONNXGPUStream(Ort::SessionOptions &, int32_t, int32_t *)
virtual void StreamWaitForEvents(int32_t stream, deviceEvent *evList, int32_t nEvents=1)
virtual size_t GPUMemCpy(void *dst, const void *src, size_t size, int32_t stream, int32_t toGPU, deviceEvent *ev=nullptr, deviceEvent *evList=nullptr, int32_t nEvents=1)
size_t TransferMemoryResourceLinkToGPU(int16_t res, int32_t stream=-1, deviceEvent *ev=nullptr, deviceEvent *evList=nullptr, int32_t nEvents=1)
void runKernelBackend(const krnlSetupTime &_xyz, const Args &... args)
virtual size_t GPUMemCpyAlways(bool onGpu, void *dst, const void *src, size_t size, int32_t stream, int32_t toGPU, deviceEvent *ev=nullptr, deviceEvent *evList=nullptr, int32_t nEvents=1)
static constexpr krnlRunRange krnlRunRangeNone
virtual bool IsEventDone(deviceEvent *evList, int32_t nEvents=1)
size_t TransferMemoryResourceLinkToHost(int16_t res, int32_t stream=-1, deviceEvent *ev=nullptr, deviceEvent *evList=nullptr, int32_t nEvents=1)
size_t TransferMemoryResourceToHost(GPUMemoryResource *res, int32_t stream=-1, deviceEvent *ev=nullptr, deviceEvent *evList=nullptr, int32_t nEvents=1)
virtual void SynchronizeStream(int32_t stream)
virtual void SynchronizeEvents(deviceEvent *evList, int32_t nEvents=1)
size_t TransferMemoryResourcesToHost(GPUProcessor *proc, int32_t stream=-1, bool all=false)
void UpdateParamOccupancyMap(const uint32_t *mapHost, const uint32_t *mapGPU, uint32_t occupancyTotal, uint32_t mapSize, int32_t stream=-1, deviceEvent *ev=nullptr)
size_t TransferMemoryResourceToGPU(GPUMemoryResource *res, int32_t stream=-1, deviceEvent *ev=nullptr, deviceEvent *evList=nullptr, int32_t nEvents=1)
int32_t unregisterMemoryForGPU_internal(const void *ptr) override
int32_t registerMemoryForGPU_internal(const void *ptr, size_t size) override
GPUProcessorProcessors mProcShadow
krnlProperties getKernelProperties(int gpu=-1)
size_t TransferMemoryResourcesToGPU(GPUProcessor *proc, int32_t stream=-1, bool all=false)
virtual int32_t GPUDebug(const char *state="UNKNOWN", int32_t stream=-1, bool force=false)
static constexpr krnlEvent krnlEventNone
virtual void ReleaseEvent(deviceEvent ev)
virtual void RecordMarker(deviceEvent *ev, int32_t stream)
size_t WriteToConstantMemory(size_t offset, const void *src, size_t size, int32_t stream=-1, deviceEvent *ev=nullptr) override
virtual size_t TransferMemoryInternal(GPUMemoryResource *res, int32_t stream, deviceEvent *ev, deviceEvent *evList, int32_t nEvents, bool toGPU, const void *src, void *dst)
gpu_reconstruction_kernels::deviceEvent deviceEvent
std::vector< GPUMemoryResource > mMemoryResources
static GPUReconstruction * GPUReconstruction_Create_CPU(const GPUSettingsDeviceBackend &cfg)
GLint GLsizei count
Definition glcorearb.h:399
GLsizeiptr size
Definition glcorearb.h:659
GLuint const GLchar * name
Definition glcorearb.h:781
GLint GLint GLsizei GLint GLenum GLenum type
Definition glcorearb.h:275
GLenum GLenum dst
Definition glcorearb.h:1767
GLintptr offset
Definition glcorearb.h:660
GLuint GLuint stream
Definition glcorearb.h:1806
const int nEvents
Definition test_Fifo.cxx:27
std::vector< int > row