HLT/TPCLib/tracking-ca/AliHLTTPCCAGPUTrackerNVCC.cu

   1 // **************************************************************************
   2 // This file is property of and copyright by the ALICE HLT Project          *
   3 // ALICE Experiment at CERN, All rights reserved.                           *
   4 //                                                                          *
   5 // Primary Authors: Sergey Gorbunov <sergey.gorbunov@kip.uni-heidelberg.de> *
   6 //                  Ivan Kisel <kisel@kip.uni-heidelberg.de>                *
   7 //                                      David Rohr <drohr@kip.uni-heidelberg.de>                                *
   8 //                  for The ALICE HLT Project.                              *
   9 //                                                                          *
  10 // Permission to use, copy, modify and distribute this software and its     *
  11 // documentation strictly for non-commercial purposes is hereby granted     *
  12 // without fee, provided that the above copyright notice appears in all     *
  13 // copies and that both the copyright notice and this permission notice     *
  14 // appear in the supporting documentation. The authors make no claims       *
  15 // about the suitability of this software for any purpose. It is            *
  16 // provided "as is" without express or implied warranty.                    *
  17 //                                                                          *
  18 //***************************************************************************
  19
  20 #include "AliHLTTPCCAGPUTracker.h"
  21
  22 #ifdef BUILD_GPU
  23
  24 #include <cuda.h>
  25 #include <sys/syscall.h>
  26
  27 #include "AliHLTTPCCADef.h"
  28 #include "AliHLTTPCCAGPUConfig.h"
  29
  30 #include <sm_11_atomic_functions.h>
  31 #include <sm_12_atomic_functions.h>
  32
  33 #include <iostream>
  34
  35 //Disable assertions since they produce errors in GPU Code
  36 #ifdef assert
  37 #undef assert
  38 #endif
  39 #define assert(param)
  40
  41 __constant__ float4 gAliHLTTPCCATracker[HLTCA_GPU_TRACKER_CONSTANT_MEM / sizeof( float4 )];
  42 #ifdef HLTCA_GPU_TEXTURE_FETCH
  43 texture<ushort2, 1, cudaReadModeElementType> gAliTexRefu2;
  44 texture<unsigned short, 1, cudaReadModeElementType> gAliTexRefu;
  45 texture<signed short, 1, cudaReadModeElementType> gAliTexRefs;
  46 #endif
  47
  48 //Include CXX Files, GPUd() macro will then produce CUDA device code out of the tracker source code
  49 #include "AliHLTTPCCATrackParam.cxx"
  50 #include "AliHLTTPCCATrack.cxx"
  51
  52 #include "AliHLTTPCCAHitArea.cxx"
  53 #include "AliHLTTPCCAGrid.cxx"
  54 #include "AliHLTTPCCARow.cxx"
  55 #include "AliHLTTPCCAParam.cxx"
  56 #include "AliHLTTPCCATracker.cxx"
  57
  58 #include "AliHLTTPCCAProcess.h"
  59
  60 #include "AliHLTTPCCATrackletSelector.cxx"
  61 #include "AliHLTTPCCANeighboursFinder.cxx"
  62 #include "AliHLTTPCCANeighboursCleaner.cxx"
  63 #include "AliHLTTPCCAStartHitsFinder.cxx"
  64 #include "AliHLTTPCCAStartHitsSorter.cxx"
  65 #include "AliHLTTPCCATrackletConstructor.cxx"
  66 #include "AliHLTTPCCASliceOutput.cxx"
  67
  68 #include "MemoryAssignmentHelpers.h"
  69
  70 #ifndef HLTCA_STANDALONE
  71 #include "AliHLTDefinitions.h"
  72 #include "AliHLTSystem.h"
  73 #endif
  74
  75 ClassImp( AliHLTTPCCAGPUTracker )
  76
  77 bool AliHLTTPCCAGPUTracker::fgGPUUsed = false;
  78
  79 int AliHLTTPCCAGPUTracker::InitGPU(int sliceCount, int forceDeviceID)
  80 {
  81         //Find best CUDA device, initialize and allocate memory
  82
  83         if (fgGPUUsed)
  84         {
  85             HLTWarning("CUDA already used by another AliHLTTPCCAGPUTracker running in same process");
  86             return(1);
  87         }
  88         fgGPUUsed = 1;
  89         fThreadId = (int) syscall (SYS_gettid);
  90
  91         cudaDeviceProp fCudaDeviceProp;
  92
  93         fGPUMemSize = HLTCA_GPU_ROWS_MEMORY + HLTCA_GPU_COMMON_MEMORY + sliceCount * (HLTCA_GPU_SLICE_DATA_MEMORY + HLTCA_GPU_GLOBAL_MEMORY);
  94
  95 #ifndef CUDA_DEVICE_EMULATION
  96         int count, bestDevice = -1;
  97         long long int bestDeviceSpeed = 0, deviceSpeed;
  98         if (CudaFailedMsg(cudaGetDeviceCount(&count)))
  99         {
 100                 HLTError("Error getting CUDA Device Count");
 101                 fgGPUUsed = 0;
 102                 return(1);
 103         }
 104         if (fDebugLevel >= 2) HLTInfo("Available CUDA devices:");
 105         for (int i = 0;i < count;i++)
 106         {
 107                 unsigned int free, total;
 108                 cuInit(0);
 109                 CUdevice tmpDevice;
 110                 cuDeviceGet(&tmpDevice, i);
 111                 CUcontext tmpContext;
 112                 cuCtxCreate(&tmpContext, 0, tmpDevice);
 113                 if(cuMemGetInfo(&free, &total)) std::cout << "Error\n";
 114                 cuCtxDestroy(tmpContext);
 115                 CudaFailedMsg(cudaGetDeviceProperties(&fCudaDeviceProp, i));
 116
 117                 int deviceOK = fCudaDeviceProp.major < 9 && !(fCudaDeviceProp.major < 1 || (fCudaDeviceProp.major == 1 && fCudaDeviceProp.minor < 2)) && free >= fGPUMemSize;
 118
 119                 if (fDebugLevel >= 2) HLTInfo("%s%2d: %s (Rev: %d.%d - Mem Avail %d / %d)%s", deviceOK ? " " : "[", i, fCudaDeviceProp.name, fCudaDeviceProp.major, fCudaDeviceProp.minor, free, fCudaDeviceProp.totalGlobalMem, deviceOK ? "" : " ]");
 120                 deviceSpeed = (long long int) fCudaDeviceProp.multiProcessorCount * (long long int) fCudaDeviceProp.clockRate * (long long int) fCudaDeviceProp.warpSize * (long long int) free;
 121                 if (deviceOK && deviceSpeed > bestDeviceSpeed)
 122                 {
 123                         bestDevice = i;
 124                         bestDeviceSpeed = deviceSpeed;
 125                 }
 126         }
 127         if (bestDevice == -1)
 128         {
 129                 HLTWarning("No CUDA Device available, aborting CUDA Initialisation");
 130                 fgGPUUsed = 0;
 131                 return(1);
 132         }
 133
 134   int cudaDevice;
 135   if (forceDeviceID == -1)
 136           cudaDevice = bestDevice;
 137   else
 138           cudaDevice = forceDeviceID;
 139 #else
 140         int cudaDevice = 0;
 141 #endif
 142
 143   cudaGetDeviceProperties(&fCudaDeviceProp ,cudaDevice );
 144
 145   if (fDebugLevel >= 1)
 146   {
 147           HLTInfo("Using CUDA Device %s with Properties:", fCudaDeviceProp.name);
 148           HLTInfo("totalGlobalMem = %d", fCudaDeviceProp.totalGlobalMem);
 149           HLTInfo("sharedMemPerBlock = %d", fCudaDeviceProp.sharedMemPerBlock);
 150           HLTInfo("regsPerBlock = %d", fCudaDeviceProp.regsPerBlock);
 151           HLTInfo("warpSize = %d", fCudaDeviceProp.warpSize);
 152           HLTInfo("memPitch = %d", fCudaDeviceProp.memPitch);
 153           HLTInfo("maxThreadsPerBlock = %d", fCudaDeviceProp.maxThreadsPerBlock);
 154           HLTInfo("maxThreadsDim = %d %d %d", fCudaDeviceProp.maxThreadsDim[0], fCudaDeviceProp.maxThreadsDim[1], fCudaDeviceProp.maxThreadsDim[2]);
 155           HLTInfo("maxGridSize = %d %d %d", fCudaDeviceProp.maxGridSize[0], fCudaDeviceProp.maxGridSize[1], fCudaDeviceProp.maxGridSize[2]);
 156           HLTInfo("totalConstMem = %d", fCudaDeviceProp.totalConstMem);
 157           HLTInfo("major = %d", fCudaDeviceProp.major);
 158           HLTInfo("minor = %d", fCudaDeviceProp.minor);
 159           HLTInfo("clockRate %d= ", fCudaDeviceProp.clockRate);
 160           HLTInfo("textureAlignment %d= ", fCudaDeviceProp.textureAlignment);
 161   }
 162
 163   if (fCudaDeviceProp.major < 1 || (fCudaDeviceProp.major == 1 && fCudaDeviceProp.minor < 2))
 164   {
 165         HLTError( "Unsupported CUDA Device" );
 166         fgGPUUsed = 0;
 167         return(1);
 168   }
 169
 170   if (sizeof(AliHLTTPCCATracker) * sliceCount > HLTCA_GPU_TRACKER_OBJECT_MEMORY)
 171   {
 172           HLTError("Insufficiant Tracker Object Memory");
 173           fgGPUUsed = 0;
 174           return(1);
 175   }
 176
 177   if (CudaFailedMsg(cudaSetDevice(cudaDevice)))
 178   {
 179           HLTError("Could not set CUDA Device!");
 180           fgGPUUsed = 0;
 181           return(1);
 182   }
 183
 184   if (fgkNSlices * AliHLTTPCCATracker::CommonMemorySize() > HLTCA_GPU_COMMON_MEMORY)
 185   {
 186           HLTError("Insufficiant Common Memory");
 187           cudaThreadExit();
 188           fgGPUUsed = 0;
 189           return(1);
 190   }
 191
 192   if (fgkNSlices * (HLTCA_ROW_COUNT + 1) * sizeof(AliHLTTPCCARow) > HLTCA_GPU_ROWS_MEMORY)
 193   {
 194           HLTError("Insufficiant Row Memory");
 195           cudaThreadExit();
 196           fgGPUUsed = 0;
 197           return(1);
 198   }
 199
 200   if (fGPUMemSize > fCudaDeviceProp.totalGlobalMem || CudaFailedMsg(cudaMalloc(&fGPUMemory, (size_t) fGPUMemSize)))
 201   {
 202           HLTError("CUDA Memory Allocation Error");
 203           cudaThreadExit();
 204           fgGPUUsed = 0;
 205           return(1);
 206   }
 207   if (fDebugLevel >= 1) HLTInfo("GPU Memory used: %d", (int) fGPUMemSize);
 208   int hostMemSize = HLTCA_GPU_ROWS_MEMORY + HLTCA_GPU_COMMON_MEMORY + sliceCount * (HLTCA_GPU_SLICE_DATA_MEMORY + HLTCA_GPU_TRACKS_MEMORY) + HLTCA_GPU_TRACKER_OBJECT_MEMORY;
 209   if (CudaFailedMsg(cudaMallocHost(&fHostLockedMemory, hostMemSize)))
 210   {
 211           cudaFree(fGPUMemory);
 212           cudaThreadExit();
 213           HLTError("Error allocating Page Locked Host Memory");
 214           fgGPUUsed = 0;
 215           return(1);
 216   }
 217   if (fDebugLevel >= 1) HLTInfo("Host Memory used: %d", hostMemSize);
 218
 219   if (fDebugLevel >= 1)
 220   {
 221           CudaFailedMsg(cudaMemset(fGPUMemory, 143, (size_t) fGPUMemSize));
 222   }
 223
 224   fSliceCount = sliceCount;
 225   //Don't run constructor / destructor here, this will be just local memcopy of Tracker in GPU Memory
 226   fGpuTracker = (AliHLTTPCCATracker*) TrackerMemory(fHostLockedMemory, 0);
 227
 228   for (int i = 0;i < fgkNSlices;i++)
 229   {
 230     fSlaveTrackers[i].SetGPUTracker();
 231         fSlaveTrackers[i].SetGPUTrackerCommonMemory((char*) CommonMemory(fHostLockedMemory, i));
 232         fSlaveTrackers[i].SetGPUSliceDataMemory(SliceDataMemory(fHostLockedMemory, i), RowMemory(fHostLockedMemory, i));
 233   }
 234
 235   fpCudaStreams = malloc(CAMath::Max(3, fSliceCount) * sizeof(cudaStream_t));
 236   cudaStream_t* const cudaStreams = (cudaStream_t*) fpCudaStreams;
 237   for (int i = 0;i < CAMath::Max(3, fSliceCount);i++)
 238   {
 239         if (CudaFailedMsg(cudaStreamCreate(&cudaStreams[i])))
 240         {
 241             cudaFree(fGPUMemory);
 242             cudaFreeHost(fHostLockedMemory);
 243             cudaThreadExit();
 244             HLTError("Error creating CUDA Stream");
 245             fgGPUUsed = 0;
 246             return(1);
 247         }
 248   }
 249
 250   fCudaInitialized = 1;
 251   HLTImportant("CUDA Initialisation successfull (Device %d: %s, Thread %dd)", cudaDevice, fCudaDeviceProp.name, fThreadId);
 252
 253 #if defined(HLTCA_STANDALONE) & !defined(CUDA_DEVICE_EMULATION)
 254   if (fDebugLevel < 2)
 255   {
 256           //Do one initial run for Benchmark reasons
 257           const int useDebugLevel = fDebugLevel;
 258           fDebugLevel = 0;
 259           AliHLTTPCCAClusterData* tmpCluster = new AliHLTTPCCAClusterData[sliceCount];
 260
 261           std::ifstream fin;
 262
 263           AliHLTTPCCAParam tmpParam;
 264           AliHLTTPCCASliceOutput::outputControlStruct tmpOutputControl;
 265
 266           fin.open("events/settings.dump");
 267           int tmpCount;
 268           fin >> tmpCount;
 269           for (int i = 0;i < sliceCount;i++)
 270           {
 271                 fSlaveTrackers[i].SetOutputControl(&tmpOutputControl);
 272                 tmpParam.ReadSettings(fin);
 273                 InitializeSliceParam(i, tmpParam);
 274           }
 275           fin.close();
 276
 277           fin.open("eventspbpbc/event.0.dump", std::ifstream::binary);
 278           for (int i = 0;i < sliceCount;i++)
 279           {
 280                 tmpCluster[i].StartReading(i, 0);
 281                 tmpCluster[i].ReadEvent(fin);
 282                 tmpCluster[i].FinishReading();
 283           }
 284           fin.close();
 285
 286           AliHLTTPCCASliceOutput **tmpOutput = new AliHLTTPCCASliceOutput*[sliceCount];
 287           memset(tmpOutput, 0, sliceCount * sizeof(AliHLTTPCCASliceOutput*));
 288
 289           Reconstruct(tmpOutput, tmpCluster, 0, sliceCount);
 290           for (int i = 0;i < sliceCount;i++)
 291           {
 292                   free(tmpOutput[i]);
 293                   tmpOutput[i] = NULL;
 294                   fSlaveTrackers[i].SetOutputControl(NULL);
 295           }
 296           delete[] tmpOutput;
 297           delete[] tmpCluster;
 298           fDebugLevel = useDebugLevel;
 299   }
 300 #endif
 301   return(0);
 302 }
 303
 304 template <class T> inline T* AliHLTTPCCAGPUTracker::alignPointer(T* ptr, int alignment)
 305 {
 306         //Macro to align Pointers.
 307         //Will align to start at 1 MB segments, this should be consistent with every alignment in the tracker
 308         //(As long as every single data structure is <= 1 MB)
 309
 310         size_t adr = (size_t) ptr;
 311         if (adr % alignment)
 312         {
 313                 adr += alignment - (adr % alignment);
 314         }
 315         return((T*) adr);
 316 }
 317
 318 bool AliHLTTPCCAGPUTracker::CudaFailedMsg(cudaError_t error)
 319 {
 320         //Check for CUDA Error and in the case of an error display the corresponding error string
 321         if (error == cudaSuccess) return(false);
 322         HLTWarning("CUDA Error: %d / %s", error, cudaGetErrorString(error));
 323         return(true);
 324 }
 325
 326 int AliHLTTPCCAGPUTracker::CUDASync(char* state)
 327 {
 328         //Wait for CUDA-Kernel to finish and check for CUDA errors afterwards
 329
 330         if (fDebugLevel == 0) return(0);
 331         cudaError cuErr;
 332         cuErr = cudaGetLastError();
 333         if (cuErr != cudaSuccess)
 334         {
 335                 HLTError("Cuda Error %s while invoking kernel (%s)", cudaGetErrorString(cuErr), state);
 336                 return(1);
 337         }
 338         if (CudaFailedMsg(cudaThreadSynchronize()))
 339         {
 340                 HLTError("CUDA Error while synchronizing (%s)", state);
 341                 return(1);
 342         }
 343         if (fDebugLevel >= 3) HLTInfo("CUDA Sync Done");
 344         return(0);
 345 }
 346
 347 void AliHLTTPCCAGPUTracker::SetDebugLevel(const int dwLevel, std::ostream* const NewOutFile)
 348 {
 349         //Set Debug Level and Debug output File if applicable
 350         fDebugLevel = dwLevel;
 351         if (NewOutFile) fOutFile = NewOutFile;
 352 }
 353
 354 int AliHLTTPCCAGPUTracker::SetGPUTrackerOption(char* OptionName, int /*OptionValue*/)
 355 {
 356         //Set a specific GPU Tracker Option
 357         {
 358                 HLTError("Unknown Option: %s", OptionName);
 359                 return(1);
 360         }
 361         //No Options used at the moment
 362         //return(0);
 363 }
 364
 365 #ifdef HLTCA_STANDALONE
 366 void AliHLTTPCCAGPUTracker::StandalonePerfTime(int iSlice, int i)
 367 {
 368   //Run Performance Query for timer i of slice iSlice
 369   if (fDebugLevel >= 1)
 370   {
 371           AliHLTTPCCAStandaloneFramework::StandaloneQueryTime( fSlaveTrackers[iSlice].PerfTimer(i));
 372   }
 373 }
 374 #else
 375 void AliHLTTPCCAGPUTracker::StandalonePerfTime(int /*iSlice*/, int /*i*/) {}
 376 #endif
 377
 378 void AliHLTTPCCAGPUTracker::DumpRowBlocks(AliHLTTPCCATracker* tracker, int iSlice, bool check)
 379 {
 380         //Dump Rowblocks to File
 381         if (fDebugLevel >= 4)
 382         {
 383                 *fOutFile << "RowBlock Tracklets" << std::endl;
 384
 385                 int4* rowBlockPos = (int4*) malloc(sizeof(int4) * (tracker[iSlice].Param().NRows() / HLTCA_GPU_SCHED_ROW_STEP + 1) * 2);
 386                 int* rowBlockTracklets = (int*) malloc(sizeof(int) * (tracker[iSlice].Param().NRows() / HLTCA_GPU_SCHED_ROW_STEP + 1) * HLTCA_GPU_MAX_TRACKLETS * 2);
 387                 uint2* blockStartingTracklet = (uint2*) malloc(sizeof(uint2) * HLTCA_GPU_BLOCK_COUNT);
 388                 CudaFailedMsg(cudaMemcpy(rowBlockPos, fGpuTracker[iSlice].RowBlockPos(), sizeof(int4) * (tracker[iSlice].Param().NRows() / HLTCA_GPU_SCHED_ROW_STEP + 1) * 2, cudaMemcpyDeviceToHost));
 389                 CudaFailedMsg(cudaMemcpy(rowBlockTracklets, fGpuTracker[iSlice].RowBlockTracklets(), sizeof(int) * (tracker[iSlice].Param().NRows() / HLTCA_GPU_SCHED_ROW_STEP + 1) * HLTCA_GPU_MAX_TRACKLETS * 2, cudaMemcpyDeviceToHost));
 390                 CudaFailedMsg(cudaMemcpy(blockStartingTracklet, fGpuTracker[iSlice].BlockStartingTracklet(), sizeof(uint2) * HLTCA_GPU_BLOCK_COUNT, cudaMemcpyDeviceToHost));
 391                 CudaFailedMsg(cudaMemcpy(tracker[iSlice].CommonMemory(), fGpuTracker[iSlice].CommonMemory(), fGpuTracker[iSlice].CommonMemorySize(), cudaMemcpyDeviceToHost));
 392
 393                 int k = tracker[iSlice].GPUParameters()->fScheduleFirstDynamicTracklet;
 394                 for (int i = 0; i < tracker[iSlice].Param().NRows() / HLTCA_GPU_SCHED_ROW_STEP + 1;i++)
 395                 {
 396                         *fOutFile << "Rowblock: " << i << ", up " << rowBlockPos[i].y << "/" << rowBlockPos[i].x << ", down " <<
 397                                 rowBlockPos[tracker[iSlice].Param().NRows() / HLTCA_GPU_SCHED_ROW_STEP + 1 + i].y << "/" << rowBlockPos[tracker[iSlice].Param().NRows() / HLTCA_GPU_SCHED_ROW_STEP + 1 + i].x << std::endl << "Phase 1: ";
 398                         for (int j = 0;j < rowBlockPos[i].x;j++)
 399                         {
 400                                 //Use Tracker Object to calculate Offset instead of fGpuTracker, since *fNTracklets of fGpuTracker points to GPU Mem!
 401                                 *fOutFile << rowBlockTracklets[(tracker[iSlice].RowBlockTracklets(0, i) - tracker[iSlice].RowBlockTracklets(0, 0)) + j] << ", ";
 402                                 if (check && rowBlockTracklets[(tracker[iSlice].RowBlockTracklets(0, i) - tracker[iSlice].RowBlockTracklets(0, 0)) + j] != k)
 403                                 {
 404                                         HLTError("Wrong starting Row Block %d, entry %d, is %d, should be %d", i, j, rowBlockTracklets[(tracker[iSlice].RowBlockTracklets(0, i) - tracker[iSlice].RowBlockTracklets(0, 0)) + j], k);
 405                                 }
 406                                 k++;
 407                                 if (rowBlockTracklets[(tracker[iSlice].RowBlockTracklets(0, i) - tracker[iSlice].RowBlockTracklets(0, 0)) + j] == -1)
 408                                 {
 409                                         HLTError("Error, -1 Tracklet found");
 410                                 }
 411                         }
 412                         *fOutFile << std::endl << "Phase 2: ";
 413                         for (int j = 0;j < rowBlockPos[tracker[iSlice].Param().NRows() / HLTCA_GPU_SCHED_ROW_STEP + 1 + i].x;j++)
 414                         {
 415                                 *fOutFile << rowBlockTracklets[(tracker[iSlice].RowBlockTracklets(1, i) - tracker[iSlice].RowBlockTracklets(0, 0)) + j] << ", ";
 416                         }
 417                         *fOutFile << std::endl;
 418                 }
 419
 420                 if (check)
 421                 {
 422                         *fOutFile << "Starting Threads: (First Dynamic: " << tracker[iSlice].GPUParameters()->fScheduleFirstDynamicTracklet << ")" << std::endl;
 423                         for (int i = 0;i < HLTCA_GPU_BLOCK_COUNT;i++)
 424                         {
 425                                 *fOutFile << i << ": " << blockStartingTracklet[i].x << " - " << blockStartingTracklet[i].y << std::endl;
 426                         }
 427                 }
 428
 429                 free(rowBlockPos);
 430                 free(rowBlockTracklets);
 431                 free(blockStartingTracklet);
 432         }
 433 }
 434
 435 __global__ void PreInitRowBlocks(int4* const RowBlockPos, int* const RowBlockTracklets, int* const SliceDataHitWeights, int nSliceDataHits)
 436 {
 437         //Initialize GPU RowBlocks and HitWeights
 438         int4* const rowBlockTracklets4 = (int4*) RowBlockTracklets;
 439         int4* const sliceDataHitWeights4 = (int4*) SliceDataHitWeights;
 440         const int stride = blockDim.x * gridDim.x;
 441         int4 i0, i1;
 442         i0.x = i0.y = i0.z = i0.w = 0;
 443         i1.x = i1.y = i1.z = i1.w = -1;
 444         for (int i = blockIdx.x * blockDim.x + threadIdx.x;i < sizeof(int4) * 2 * (HLTCA_ROW_COUNT / HLTCA_GPU_SCHED_ROW_STEP + 1) / sizeof(int4);i += stride)
 445                 RowBlockPos[i] = i0;
 446         for (int i = blockIdx.x * blockDim.x + threadIdx.x;i < sizeof(int) * (HLTCA_ROW_COUNT / HLTCA_GPU_SCHED_ROW_STEP + 1) * HLTCA_GPU_MAX_TRACKLETS * 2 / sizeof(int4);i += stride)
 447                 rowBlockTracklets4[i] = i1;
 448         for (int i = blockIdx.x * blockDim.x + threadIdx.x;i < nSliceDataHits * sizeof(int) / sizeof(int4);i += stride)
 449                 sliceDataHitWeights4[i] = i0;
 450 }
 451
 452 int AliHLTTPCCAGPUTracker::Reconstruct(AliHLTTPCCASliceOutput** pOutput, AliHLTTPCCAClusterData* pClusterData, int firstSlice, int sliceCountLocal)
 453 {
 454         //Primary reconstruction function
 455         cudaStream_t* const cudaStreams = (cudaStream_t*) fpCudaStreams;
 456
 457         if (sliceCountLocal == -1) sliceCountLocal = fSliceCount;
 458
 459         if (!fCudaInitialized)
 460         {
 461             HLTError("GPUTracker not initialized");
 462             return(1);
 463         }
 464         if (sliceCountLocal > fSliceCount)
 465         {
 466             HLTError("GPU Tracker was initialized to run with %d slices but was called to process %d slices", fSliceCount, sliceCountLocal);
 467             return(1);
 468         }
 469         if (fThreadId != (int) syscall (SYS_gettid))
 470         {
 471             HLTError("GPUTracker context was initialized by different thread, Initializing Thread: %d, Processing Thread: %d", fThreadId, (int) syscall (SYS_gettid));
 472             return(1);
 473         }
 474
 475         if (fDebugLevel >= 2) HLTInfo("Running GPU Tracker (Slices %d to %d)", fSlaveTrackers[firstSlice].Param().ISlice(), fSlaveTrackers[firstSlice].Param().ISlice() + sliceCountLocal);
 476
 477         if (sliceCountLocal * sizeof(AliHLTTPCCATracker) > HLTCA_GPU_TRACKER_CONSTANT_MEM)
 478         {
 479                 HLTError("Insuffissant constant memory (Required %d, Available %d, Tracker %d, Param %d, SliceData %d)", sliceCountLocal * (int) sizeof(AliHLTTPCCATracker), (int) HLTCA_GPU_TRACKER_CONSTANT_MEM, (int) sizeof(AliHLTTPCCATracker), (int) sizeof(AliHLTTPCCAParam), (int) sizeof(AliHLTTPCCASliceData));
 480                 return(1);
 481         }
 482
 483         if (fDebugLevel >= 4)
 484         {
 485                 for (int iSlice = 0;iSlice < sliceCountLocal;iSlice++)
 486                 {
 487                         *fOutFile << std::endl << std::endl << "Slice: " << fSlaveTrackers[firstSlice + iSlice].Param().ISlice() << std::endl;
 488                 }
 489         }
 490
 491         memcpy(fGpuTracker, &fSlaveTrackers[firstSlice], sizeof(AliHLTTPCCATracker) * sliceCountLocal);
 492
 493         if (fDebugLevel >= 3) HLTInfo("Allocating GPU Tracker memory and initializing constants");
 494
 495 #ifdef HLTCA_GPU_TIME_PROFILE
 496         unsigned __int64 a, b, c, d;
 497         AliHLTTPCCAStandaloneFramework::StandaloneQueryFreq(&c);
 498         AliHLTTPCCAStandaloneFramework::StandaloneQueryTime(&d);
 499 #endif
 500
 501         for (int iSlice = 0;iSlice < sliceCountLocal;iSlice++)
 502         {
 503                 //Make this a GPU Tracker
 504                 fGpuTracker[iSlice].SetGPUTracker();
 505                 fGpuTracker[iSlice].SetGPUTrackerCommonMemory((char*) CommonMemory(fGPUMemory, iSlice));
 506                 fGpuTracker[iSlice].SetGPUSliceDataMemory(SliceDataMemory(fGPUMemory, iSlice), RowMemory(fGPUMemory, iSlice));
 507                 fGpuTracker[iSlice].SetPointersSliceData(&pClusterData[iSlice], false);
 508
 509                 //Set Pointers to GPU Memory
 510                 char* tmpMem = (char*) GlobalMemory(fGPUMemory, iSlice);
 511
 512                 if (fDebugLevel >= 3) HLTInfo("Initialising GPU Hits Memory");
 513                 tmpMem = fGpuTracker[iSlice].SetGPUTrackerHitsMemory(tmpMem, pClusterData[iSlice].NumberOfClusters());
 514                 tmpMem = alignPointer(tmpMem, 1024 * 1024);
 515
 516                 if (fDebugLevel >= 3) HLTInfo("Initialising GPU Tracklet Memory");
 517                 tmpMem = fGpuTracker[iSlice].SetGPUTrackerTrackletsMemory(tmpMem, HLTCA_GPU_MAX_TRACKLETS /* *fSlaveTrackers[firstSlice + iSlice].NTracklets()*/);
 518                 tmpMem = alignPointer(tmpMem, 1024 * 1024);
 519
 520                 if (fDebugLevel >= 3) HLTInfo("Initialising GPU Track Memory");
 521                 tmpMem = fGpuTracker[iSlice].SetGPUTrackerTracksMemory(tmpMem, HLTCA_GPU_MAX_TRACKS /* *fSlaveTrackers[firstSlice + iSlice].NTracklets()*/, pClusterData[iSlice].NumberOfClusters());
 522                 tmpMem = alignPointer(tmpMem, 1024 * 1024);
 523
 524                 if (fGpuTracker[iSlice].TrackMemorySize() >= HLTCA_GPU_TRACKS_MEMORY)
 525                 {
 526                         HLTError("Insufficiant Track Memory");
 527                         return(1);
 528                 }
 529
 530                 if (tmpMem - (char*) GlobalMemory(fGPUMemory, iSlice) > HLTCA_GPU_GLOBAL_MEMORY)
 531                 {
 532                         HLTError("Insufficiant Global Memory");
 533                         return(1);
 534                 }
 535
 536                 //Initialize Startup Constants
 537                 *fSlaveTrackers[firstSlice + iSlice].NTracklets() = 0;
 538                 *fSlaveTrackers[firstSlice + iSlice].NTracks() = 0;
 539                 *fSlaveTrackers[firstSlice + iSlice].NTrackHits() = 0;
 540                 fGpuTracker[iSlice].GPUParametersConst()->fGPUFixedBlockCount = HLTCA_GPU_BLOCK_COUNT * (iSlice + 1) / sliceCountLocal - HLTCA_GPU_BLOCK_COUNT * (iSlice) / sliceCountLocal;
 541                 if (fDebugLevel >= 3) HLTInfo("Blocks for Slice %d: %d", iSlice, fGpuTracker[iSlice].GPUParametersConst()->fGPUFixedBlockCount);
 542                 fGpuTracker[iSlice].GPUParametersConst()->fGPUiSlice = iSlice;
 543                 fGpuTracker[iSlice].GPUParametersConst()->fGPUnSlices = sliceCountLocal;
 544                 fSlaveTrackers[firstSlice + iSlice].GPUParameters()->fGPUError = 0;
 545                 fGpuTracker[iSlice].SetGPUTextureBase(fGpuTracker[0].Data().Memory());
 546         }
 547
 548 #ifdef HLTCA_GPU_TEXTURE_FETCH
 549                 cudaChannelFormatDesc channelDescu2 = cudaCreateChannelDesc<ushort2>();
 550                 size_t offset;
 551                 if (CudaFailedMsg(cudaBindTexture(&offset, &gAliTexRefu2, fGpuTracker[0].Data().Memory(), &channelDescu2, sliceCountLocal * HLTCA_GPU_SLICE_DATA_MEMORY)) || offset)
 552                 {
 553                         HLTError("Error binding CUDA Texture ushort2 (Offset %d)", (int) offset);
 554                         return(1);
 555                 }
 556                 cudaChannelFormatDesc channelDescu = cudaCreateChannelDesc<unsigned short>();
 557                 if (CudaFailedMsg(cudaBindTexture(&offset, &gAliTexRefu, fGpuTracker[0].Data().Memory(), &channelDescu, sliceCountLocal * HLTCA_GPU_SLICE_DATA_MEMORY)) || offset)
 558                 {
 559                         HLTError("Error binding CUDA Texture ushort (Offset %d)", (int) offset);
 560                         return(1);
 561                 }
 562                 cudaChannelFormatDesc channelDescs = cudaCreateChannelDesc<signed short>();
 563                 if (CudaFailedMsg(cudaBindTexture(&offset, &gAliTexRefs, fGpuTracker[0].Data().Memory(), &channelDescs, sliceCountLocal * HLTCA_GPU_SLICE_DATA_MEMORY)) || offset)
 564                 {
 565                         HLTError("Error binding CUDA Texture short (Offset %d)", (int) offset);
 566                         return(1);
 567                 }
 568 #endif
 569
 570         //Copy Tracker Object to GPU Memory
 571         if (fDebugLevel >= 3) HLTInfo("Copying Tracker objects to GPU");
 572 #ifdef HLTCA_GPU_TRACKLET_CONSTRUCTOR_DO_PROFILE
 573         if (CudaFailedMsg(cudaMalloc(&fGpuTracker[0].fStageAtSync, 100000000))) return(1);
 574         CudaFailedMsg(cudaMemset(fGpuTracker[0].fStageAtSync, 0, 100000000));
 575 #endif
 576         CudaFailedMsg(cudaMemcpyToSymbolAsync(gAliHLTTPCCATracker, fGpuTracker, sizeof(AliHLTTPCCATracker) * sliceCountLocal, 0, cudaMemcpyHostToDevice, cudaStreams[0]));
 577
 578         for (int iSlice = 0;iSlice < sliceCountLocal;iSlice++)
 579         {
 580                 StandalonePerfTime(firstSlice + iSlice, 0);
 581
 582                 //Initialize GPU Slave Tracker
 583                 if (fDebugLevel >= 3) HLTInfo("Creating Slice Data");
 584                 fSlaveTrackers[firstSlice + iSlice].SetGPUSliceDataMemory(SliceDataMemory(fHostLockedMemory, iSlice), RowMemory(fHostLockedMemory, firstSlice + iSlice));
 585 #ifdef HLTCA_GPU_TIME_PROFILE
 586                         AliHLTTPCCAStandaloneFramework::StandaloneQueryTime(&a);
 587 #endif
 588                 fSlaveTrackers[firstSlice + iSlice].ReadEvent(&pClusterData[iSlice]);
 589 #ifdef HLTCA_GPU_TIME_PROFILE
 590                         AliHLTTPCCAStandaloneFramework::StandaloneQueryTime(&b);
 591                 printf("Read %f %f\n", ((double) b - (double) a) / (double) c, ((double) a - (double) d) / (double) c);
 592 #endif
 593                 if (fSlaveTrackers[firstSlice + iSlice].Data().MemorySize() > HLTCA_GPU_SLICE_DATA_MEMORY)
 594                 {
 595                         HLTError("Insufficiant Slice Data Memory");
 596                         return(1);
 597                 }
 598
 599                 /*if (fSlaveTrackers[firstSlice + iSlice].CheckEmptySlice())
 600                 {
 601                         if (fDebugLevel >= 3) HLTInfo("Slice Empty, not running GPU Tracker");
 602                         if (sliceCountLocal == 1)
 603                                 return(0);
 604                 }*/
 605
 606                 //Initialize temporary memory where needed
 607                 if (fDebugLevel >= 3) HLTInfo("Copying Slice Data to GPU and initializing temporary memory");
 608                 PreInitRowBlocks<<<30, 256, 0, cudaStreams[2]>>>(fGpuTracker[iSlice].RowBlockPos(), fGpuTracker[iSlice].RowBlockTracklets(), fGpuTracker[iSlice].Data().HitWeights(), fSlaveTrackers[firstSlice + iSlice].Data().NumberOfHitsPlusAlign());
 609
 610                 //Copy Data to GPU Global Memory
 611                 CudaFailedMsg(cudaMemcpyAsync(fGpuTracker[iSlice].CommonMemory(), fSlaveTrackers[firstSlice + iSlice].CommonMemory(), fSlaveTrackers[firstSlice + iSlice].CommonMemorySize(), cudaMemcpyHostToDevice, cudaStreams[iSlice & 1]));
 612                 CudaFailedMsg(cudaMemcpyAsync(fGpuTracker[iSlice].Data().Memory(), fSlaveTrackers[firstSlice + iSlice].Data().Memory(), fSlaveTrackers[firstSlice + iSlice].Data().GpuMemorySize(), cudaMemcpyHostToDevice, cudaStreams[iSlice & 1]));
 613                 CudaFailedMsg(cudaMemcpyAsync(fGpuTracker[iSlice].SliceDataRows(), fSlaveTrackers[firstSlice + iSlice].SliceDataRows(), (HLTCA_ROW_COUNT + 1) * sizeof(AliHLTTPCCARow), cudaMemcpyHostToDevice, cudaStreams[iSlice & 1]));
 614
 615                 if (fDebugLevel >= 4)
 616                 {
 617                         if (fDebugLevel >= 5) HLTInfo("Allocating Debug Output Memory");
 618                         fSlaveTrackers[firstSlice + iSlice].TrackletMemory() = reinterpret_cast<char*> ( new uint4 [ fGpuTracker[iSlice].TrackletMemorySize()/sizeof( uint4 ) + 100] );
 619                         fSlaveTrackers[firstSlice + iSlice].SetPointersTracklets( HLTCA_GPU_MAX_TRACKLETS );
 620                         fSlaveTrackers[firstSlice + iSlice].HitMemory() = reinterpret_cast<char*> ( new uint4 [ fGpuTracker[iSlice].HitMemorySize()/sizeof( uint4 ) + 100] );
 621                         fSlaveTrackers[firstSlice + iSlice].SetPointersHits( pClusterData[iSlice].NumberOfClusters() );
 622                 }
 623
 624                 if (CUDASync("Initialization")) return(1);
 625                 StandalonePerfTime(firstSlice + iSlice, 1);
 626
 627                 if (fDebugLevel >= 3) HLTInfo("Running GPU Neighbours Finder");
 628                 AliHLTTPCCAProcess<AliHLTTPCCANeighboursFinder> <<<fSlaveTrackers[firstSlice + iSlice].Param().NRows(), 256, 0, cudaStreams[iSlice & 1]>>>(iSlice);
 629
 630                 if (CUDASync("Neighbours finder")) return 1;
 631
 632                 StandalonePerfTime(firstSlice + iSlice, 2);
 633
 634                 if (fDebugLevel >= 4)
 635                 {
 636                         CudaFailedMsg(cudaMemcpy(fSlaveTrackers[firstSlice + iSlice].Data().Memory(), fGpuTracker[iSlice].Data().Memory(), fSlaveTrackers[firstSlice + iSlice].Data().GpuMemorySize(), cudaMemcpyDeviceToHost));
 637                         fSlaveTrackers[firstSlice + iSlice].DumpLinks(*fOutFile);
 638                 }
 639
 640                 if (fDebugLevel >= 3) HLTInfo("Running GPU Neighbours Cleaner");
 641                 AliHLTTPCCAProcess<AliHLTTPCCANeighboursCleaner> <<<fSlaveTrackers[firstSlice + iSlice].Param().NRows()-2, 256, 0, cudaStreams[iSlice & 1]>>>(iSlice);
 642                 if (CUDASync("Neighbours Cleaner")) return 1;
 643
 644                 StandalonePerfTime(firstSlice + iSlice, 3);
 645
 646                 if (fDebugLevel >= 4)
 647                 {
 648                         CudaFailedMsg(cudaMemcpy(fSlaveTrackers[firstSlice + iSlice].Data().Memory(), fGpuTracker[iSlice].Data().Memory(), fSlaveTrackers[firstSlice + iSlice].Data().GpuMemorySize(), cudaMemcpyDeviceToHost));
 649                         fSlaveTrackers[firstSlice + iSlice].DumpLinks(*fOutFile);
 650                 }
 651
 652                 if (fDebugLevel >= 3) HLTInfo("Running GPU Start Hits Finder");
 653                 AliHLTTPCCAProcess<AliHLTTPCCAStartHitsFinder> <<<fSlaveTrackers[firstSlice + iSlice].Param().NRows()-6, 256, 0, cudaStreams[iSlice & 1]>>>(iSlice);
 654                 if (CUDASync("Start Hits Finder")) return 1;
 655
 656                 StandalonePerfTime(firstSlice + iSlice, 4);
 657
 658                 if (fDebugLevel >= 3) HLTInfo("Running GPU Start Hits Sorter");
 659                 AliHLTTPCCAProcess<AliHLTTPCCAStartHitsSorter> <<<30, 256, 0, cudaStreams[iSlice & 1]>>>(iSlice);
 660                 if (CUDASync("Start Hits Sorter")) return 1;
 661
 662                 StandalonePerfTime(firstSlice + iSlice, 5);
 663
 664                 if (fDebugLevel >= 2)
 665                 {
 666                         CudaFailedMsg(cudaMemcpy(fSlaveTrackers[firstSlice + iSlice].CommonMemory(), fGpuTracker[iSlice].CommonMemory(), fGpuTracker[iSlice].CommonMemorySize(), cudaMemcpyDeviceToHost));
 667                         if (fDebugLevel >= 3) HLTInfo("Obtaining Number of Start Hits from GPU: %d", *fSlaveTrackers[firstSlice + iSlice].NTracklets());
 668                         if (*fSlaveTrackers[firstSlice + iSlice].NTracklets() > HLTCA_GPU_MAX_TRACKLETS)
 669                         {
 670                                 HLTError("HLTCA_GPU_MAX_TRACKLETS constant insuffisant");
 671                                 return(1);
 672                         }
 673                 }
 674
 675                 if (fDebugLevel >= 4)
 676                 {
 677                         *fOutFile << "Temporary ";
 678                         CudaFailedMsg(cudaMemcpy(fSlaveTrackers[firstSlice + iSlice].TrackletStartHits(), fGpuTracker[iSlice].TrackletTmpStartHits(), pClusterData[iSlice].NumberOfClusters() * sizeof(AliHLTTPCCAHitId), cudaMemcpyDeviceToHost));
 679                         fSlaveTrackers[firstSlice + iSlice].DumpStartHits(*fOutFile);
 680                         uint3* tmpMemory = (uint3*) malloc(sizeof(uint3) * fSlaveTrackers[firstSlice + iSlice].Param().NRows());
 681                         CudaFailedMsg(cudaMemcpy(tmpMemory, fGpuTracker[iSlice].RowStartHitCountOffset(), fSlaveTrackers[firstSlice + iSlice].Param().NRows() * sizeof(uint3), cudaMemcpyDeviceToHost));
 682                         *fOutFile << "Start Hits Sort Vector:" << std::endl;
 683                         for (int i = 0;i < fSlaveTrackers[firstSlice + iSlice].Param().NRows();i++)
 684                         {
 685                                 *fOutFile << "Row: " << i << ", Len: " << tmpMemory[i].x << ", Offset: " << tmpMemory[i].y << ", New Offset: " << tmpMemory[i].z << std::endl;
 686                         }
 687                         free(tmpMemory);
 688
 689                         CudaFailedMsg(cudaMemcpy(fSlaveTrackers[firstSlice + iSlice].HitMemory(), fGpuTracker[iSlice].HitMemory(), fSlaveTrackers[firstSlice + iSlice].HitMemorySize(), cudaMemcpyDeviceToHost));
 690                         fSlaveTrackers[firstSlice + iSlice].DumpStartHits(*fOutFile);
 691                 }
 692
 693                 StandalonePerfTime(firstSlice + iSlice, 6);
 694
 695                 fSlaveTrackers[firstSlice + iSlice].SetGPUTrackerTracksMemory((char*) TracksMemory(fHostLockedMemory, iSlice), HLTCA_GPU_MAX_TRACKS, pClusterData[iSlice].NumberOfClusters());
 696         }
 697
 698         StandalonePerfTime(firstSlice, 7);
 699 #ifdef HLTCA_GPU_PREFETCHDATA
 700         for (int iSlice = 0;iSlice < sliceCountLocal;iSlice++)
 701         {
 702                 if (fSlaveTrackers[firstSlice + iSlice].Data().GPUSharedDataReq() * sizeof(ushort_v) > ALIHLTTPCCATRACKLET_CONSTRUCTOR_TEMP_MEM / 4 * sizeof(uint4))
 703                 {
 704                         HLTError("Insufficiant GPU shared Memory, required: %d, available %d", fSlaveTrackers[firstSlice + iSlice].Data().GPUSharedDataReq() * sizeof(ushort_v), ALIHLTTPCCATRACKLET_CONSTRUCTOR_TEMP_MEM / 4 * sizeof(uint4));
 705                         return(1);
 706                 }
 707                 if (fDebugLevel >= 1)
 708                 {
 709                         static int infoShown = 0;
 710                         if (!infoShown)
 711                         {
 712                                 HLTInfo("GPU Shared Memory Cache Size: %d", 2 * fSlaveTrackers[firstSlice + iSlice].Data().GPUSharedDataReq() * sizeof(ushort_v));
 713                                 infoShown = 1;
 714                         }
 715                 }
 716         }
 717 #endif
 718
 719         int nHardCollisions = 0;
 720
 721 RestartTrackletConstructor:
 722         if (fDebugLevel >= 3) HLTInfo("Initialising Tracklet Constructor Scheduler");
 723         for (int iSlice = 0;iSlice < sliceCountLocal;iSlice++)
 724         {
 725                 AliHLTTPCCATrackletConstructorInit<<<HLTCA_GPU_MAX_TRACKLETS /* *fSlaveTrackers[firstSlice + iSlice].NTracklets() */ / HLTCA_GPU_THREAD_COUNT + 1, HLTCA_GPU_THREAD_COUNT>>>(iSlice);
 726                 if (CUDASync("Tracklet Initializer")) return 1;
 727                 DumpRowBlocks(fSlaveTrackers, iSlice);
 728         }
 729
 730         if (fDebugLevel >= 3) HLTInfo("Running GPU Tracklet Constructor");
 731         AliHLTTPCCATrackletConstructorNewGPU<<<HLTCA_GPU_BLOCK_COUNT, HLTCA_GPU_THREAD_COUNT>>>();
 732         if (CUDASync("Tracklet Constructor (new)")) return 1;
 733
 734         StandalonePerfTime(firstSlice, 8);
 735
 736         if (fDebugLevel >= 4)
 737         {
 738                 for (int iSlice = 0;iSlice < sliceCountLocal;iSlice++)
 739                 {
 740                         DumpRowBlocks(&fSlaveTrackers[firstSlice], iSlice, false);
 741                         CudaFailedMsg(cudaMemcpy(fSlaveTrackers[firstSlice + iSlice].CommonMemory(), fGpuTracker[iSlice].CommonMemory(), fGpuTracker[iSlice].CommonMemorySize(), cudaMemcpyDeviceToHost));
 742                         if (fDebugLevel >= 5)
 743                         {
 744                                 HLTInfo("Obtained %d tracklets", *fSlaveTrackers[firstSlice + iSlice].NTracklets());
 745                         }
 746                         CudaFailedMsg(cudaMemcpy(fSlaveTrackers[firstSlice + iSlice].TrackletMemory(), fGpuTracker[iSlice].TrackletMemory(), fGpuTracker[iSlice].TrackletMemorySize(), cudaMemcpyDeviceToHost));
 747                         CudaFailedMsg(cudaMemcpy(fSlaveTrackers[firstSlice + iSlice].HitMemory(), fGpuTracker[iSlice].HitMemory(), fGpuTracker[iSlice].HitMemorySize(), cudaMemcpyDeviceToHost));
 748                         fSlaveTrackers[firstSlice + iSlice].DumpTrackletHits(*fOutFile);
 749                 }
 750         }
 751
 752         int runSlices = 0;
 753         for (int iSlice = 0;iSlice < sliceCountLocal;iSlice += runSlices)
 754         {
 755                 if (runSlices < HLTCA_GPU_TRACKLET_SELECTOR_SLICE_COUNT) runSlices++;
 756                 if (fDebugLevel >= 3) HLTInfo("Running HLT Tracklet selector (Slice %d to %d)", iSlice, iSlice + runSlices);
 757                 AliHLTTPCCAProcessMulti<AliHLTTPCCATrackletSelector><<<HLTCA_GPU_BLOCK_COUNT, HLTCA_GPU_THREAD_COUNT, 0, cudaStreams[iSlice]>>>(iSlice, CAMath::Min(runSlices, sliceCountLocal - iSlice));
 758         }
 759         if (CUDASync("Tracklet Selector")) return 1;
 760         StandalonePerfTime(firstSlice, 9);
 761
 762         CudaFailedMsg(cudaMemcpyAsync(fSlaveTrackers[firstSlice + 0].CommonMemory(), fGpuTracker[0].CommonMemory(), fGpuTracker[0].CommonMemorySize(), cudaMemcpyDeviceToHost, cudaStreams[0]));
 763         for (int iSliceTmp = 0;iSliceTmp <= sliceCountLocal;iSliceTmp++)
 764         {
 765                 if (iSliceTmp < sliceCountLocal)
 766                 {
 767                         int iSlice = iSliceTmp;
 768                         if (fDebugLevel >= 3) HLTInfo("Transfering Tracks from GPU to Host");
 769                         cudaStreamSynchronize(cudaStreams[iSlice]);
 770                         CudaFailedMsg(cudaMemcpyAsync(fSlaveTrackers[firstSlice + iSlice].Tracks(), fGpuTracker[iSlice].Tracks(), sizeof(AliHLTTPCCATrack) * *fSlaveTrackers[firstSlice + iSlice].NTracks(), cudaMemcpyDeviceToHost, cudaStreams[iSlice]));
 771                         CudaFailedMsg(cudaMemcpyAsync(fSlaveTrackers[firstSlice + iSlice].TrackHits(), fGpuTracker[iSlice].TrackHits(), sizeof(AliHLTTPCCAHitId) * *fSlaveTrackers[firstSlice + iSlice].NTrackHits(), cudaMemcpyDeviceToHost, cudaStreams[iSlice]));
 772                         if (iSlice + 1 < sliceCountLocal)
 773                                 CudaFailedMsg(cudaMemcpyAsync(fSlaveTrackers[firstSlice + iSlice + 1].CommonMemory(), fGpuTracker[iSlice + 1].CommonMemory(), fGpuTracker[iSlice + 1].CommonMemorySize(), cudaMemcpyDeviceToHost, cudaStreams[iSlice + 1]));
 774                 }
 775
 776                 if (iSliceTmp)
 777                 {
 778                         int iSlice = iSliceTmp - 1;
 779                         cudaStreamSynchronize(cudaStreams[iSlice]);
 780
 781                         if (fDebugLevel >= 4)
 782                         {
 783                                 CudaFailedMsg(cudaMemcpy(fSlaveTrackers[firstSlice + iSlice].Data().HitWeights(), fGpuTracker[iSlice].Data().HitWeights(), fSlaveTrackers[firstSlice + iSlice].Data().NumberOfHitsPlusAlign() * sizeof(int), cudaMemcpyDeviceToHost));
 784                                 fSlaveTrackers[firstSlice + iSlice].DumpHitWeights(*fOutFile);
 785                                 fSlaveTrackers[firstSlice + iSlice].DumpTrackHits(*fOutFile);
 786                         }
 787
 788                         if (fSlaveTrackers[firstSlice + iSlice].GPUParameters()->fGPUError)
 789                         {
 790                                 if (fSlaveTrackers[firstSlice + iSlice].GPUParameters()->fGPUError == HLTCA_GPU_ERROR_SCHEDULE_COLLISION && nHardCollisions++ < 10)
 791                                 {
 792                                         HLTWarning("Hard scheduling collision occured, rerunning Tracklet Constructor");
 793                                         for (int i = 0;i < sliceCountLocal;i++)
 794                                         {
 795                                                 cudaThreadSynchronize();
 796                                                 CudaFailedMsg(cudaMemcpy(fSlaveTrackers[firstSlice + i].CommonMemory(), fGpuTracker[i].CommonMemory(), fGpuTracker[i].CommonMemorySize(), cudaMemcpyDeviceToHost));
 797                                                 *fSlaveTrackers[firstSlice + i].NTracks() = 0;
 798                                                 *fSlaveTrackers[firstSlice + i].NTrackHits() = 0;
 799                                                 fSlaveTrackers[firstSlice + i].GPUParameters()->fGPUError = HLTCA_GPU_ERROR_NONE;
 800                                                 CudaFailedMsg(cudaMemcpy(fGpuTracker[i].CommonMemory(), fSlaveTrackers[firstSlice + i].CommonMemory(), fGpuTracker[i].CommonMemorySize(), cudaMemcpyHostToDevice));
 801                                                 PreInitRowBlocks<<<30, 256>>>(fGpuTracker[i].RowBlockPos(), fGpuTracker[i].RowBlockTracklets(), fGpuTracker[i].Data().HitWeights(), fSlaveTrackers[firstSlice + i].Data().NumberOfHitsPlusAlign());
 802                                         }
 803                                         goto RestartTrackletConstructor;
 804                                 }
 805                                 HLTError("GPU Tracker returned Error Code %d", fSlaveTrackers[firstSlice + iSlice].GPUParameters()->fGPUError);
 806                                 return(1);
 807                         }
 808                         if (fDebugLevel >= 3) HLTInfo("Tracks Transfered: %d / %d", *fSlaveTrackers[firstSlice + iSlice].NTracks(), *fSlaveTrackers[firstSlice + iSlice].NTrackHits());
 809
 810                         fSlaveTrackers[firstSlice + iSlice].SetOutput(&pOutput[iSlice]);
 811 #ifdef HLTCA_GPU_TIME_PROFILE
 812                         AliHLTTPCCAStandaloneFramework::StandaloneQueryTime(&a);
 813 #endif
 814                         fSlaveTrackers[firstSlice + iSlice].WriteOutput();
 815 #ifdef HLTCA_GPU_TIME_PROFILE
 816                         AliHLTTPCCAStandaloneFramework::StandaloneQueryTime(&b);
 817                         printf("Write %f %f\n", ((double) b - (double) a) / (double) c, ((double) a - (double) d) / (double) c);
 818 #endif
 819
 820                         if (fDebugLevel >= 4)
 821                         {
 822                                 delete[] fSlaveTrackers[firstSlice + iSlice].HitMemory();
 823                                 delete[] fSlaveTrackers[firstSlice + iSlice].TrackletMemory();
 824                         }
 825                 }
 826         }
 827
 828         StandalonePerfTime(firstSlice, 10);
 829
 830         if (fDebugLevel >= 3) HLTInfo("GPU Reconstruction finished");
 831
 832 #ifdef HLTCA_GPU_TRACKLET_CONSTRUCTOR_DO_PROFILE
 833         char* stageAtSync = (char*) malloc(100000000);
 834         CudaFailedMsg(cudaMemcpy(stageAtSync, fGpuTracker[0].fStageAtSync, 100 * 1000 * 1000, cudaMemcpyDeviceToHost));
 835         cudaFree(fGpuTracker[0].fStageAtSync);
 836
 837         FILE* fp = fopen("profile.txt", "w+");
 838         FILE* fp2 = fopen("profile.bmp", "w+b");
 839         int nEmptySync = 0, fEmpty;
 840
 841         const int bmpheight = 1000;
 842         BITMAPFILEHEADER bmpFH;
 843         BITMAPINFOHEADER bmpIH;
 844         ZeroMemory(&bmpFH, sizeof(bmpFH));
 845         ZeroMemory(&bmpIH, sizeof(bmpIH));
 846
 847         bmpFH.bfType = 19778; //"BM"
 848         bmpFH.bfSize = sizeof(bmpFH) + sizeof(bmpIH) + (HLTCA_GPU_BLOCK_COUNT * HLTCA_GPU_THREAD_COUNT / 32 * 33 - 1) * bmpheight ;
 849         bmpFH.bfOffBits = sizeof(bmpFH) + sizeof(bmpIH);
 850
 851         bmpIH.biSize = sizeof(bmpIH);
 852         bmpIH.biWidth = HLTCA_GPU_BLOCK_COUNT * HLTCA_GPU_THREAD_COUNT / 32 * 33 - 1;
 853         bmpIH.biHeight = bmpheight;
 854         bmpIH.biPlanes = 1;
 855         bmpIH.biBitCount = 32;
 856
 857         fwrite(&bmpFH, 1, sizeof(bmpFH), fp2);
 858         fwrite(&bmpIH, 1, sizeof(bmpIH), fp2);
 859
 860         for (int i = 0;i < bmpheight * HLTCA_GPU_BLOCK_COUNT * HLTCA_GPU_THREAD_COUNT;i += HLTCA_GPU_BLOCK_COUNT * HLTCA_GPU_THREAD_COUNT)
 861         {
 862                 fEmpty = 1;
 863                 for (int j = 0;j < HLTCA_GPU_BLOCK_COUNT * HLTCA_GPU_THREAD_COUNT;j++)
 864                 {
 865                         fprintf(fp, "%d\t", stageAtSync[i + j]);
 866                         int color = 0;
 867                         if (stageAtSync[i + j] == 1) color = RGB(255, 0, 0);
 868                         if (stageAtSync[i + j] == 2) color = RGB(0, 255, 0);
 869                         if (stageAtSync[i + j] == 3) color = RGB(0, 0, 255);
 870                         if (stageAtSync[i + j] == 4) color = RGB(255, 255, 0);
 871                         fwrite(&color, 1, sizeof(int), fp2);
 872                         if (j > 0 && j % 32 == 0)
 873                         {
 874                                 color = RGB(255, 255, 255);
 875                                 fwrite(&color, 1, 4, fp2);
 876                         }
 877                         if (stageAtSync[i + j]) fEmpty = 0;
 878                 }
 879                 fprintf(fp, "\n");
 880                 if (fEmpty) nEmptySync++;
 881                 else nEmptySync = 0;
 882                 //if (nEmptySync == HLTCA_GPU_SCHED_ROW_STEP + 2) break;
 883         }
 884
 885         fclose(fp);
 886         fclose(fp2);
 887         free(stageAtSync);
 888 #endif
 889
 890         return(0);
 891 }
 892
 893 int AliHLTTPCCAGPUTracker::InitializeSliceParam(int iSlice, AliHLTTPCCAParam &param)
 894 {
 895         //Initialize Slice Tracker Parameter for a slave tracker
 896         fSlaveTrackers[iSlice].Initialize(param);
 897         if (fSlaveTrackers[iSlice].Param().NRows() != HLTCA_ROW_COUNT)
 898         {
 899                 HLTError("Error, Slice Tracker %d Row Count of %d exceeds Constant of %d", iSlice, fSlaveTrackers[iSlice].Param().NRows(), HLTCA_ROW_COUNT);
 900                 return(1);
 901         }
 902         return(0);
 903 }
 904
 905 int AliHLTTPCCAGPUTracker::ExitGPU()
 906 {
 907         //Uninitialize CUDA
 908         cudaThreadSynchronize();
 909         if (fGPUMemory)
 910         {
 911                 cudaFree(fGPUMemory);
 912                 fGPUMemory = NULL;
 913         }
 914         if (fHostLockedMemory)
 915         {
 916                 for (int i = 0;i < CAMath::Max(3, fSliceCount);i++)
 917                 {
 918                         cudaStreamDestroy(((cudaStream_t*) fpCudaStreams)[i]);
 919                 }
 920                 free(fpCudaStreams);
 921                 fGpuTracker = NULL;
 922                 cudaFreeHost(fHostLockedMemory);
 923         }
 924
 925         if (CudaFailedMsg(cudaThreadExit()))
 926         {
 927                 HLTError("Could not uninitialize GPU");
 928                 return(1);
 929         }
 930         HLTInfo("CUDA Uninitialized");
 931         fgGPUUsed = false;
 932         fCudaInitialized = 0;
 933         return(0);
 934 }
 935
 936 void AliHLTTPCCAGPUTracker::SetOutputControl( AliHLTTPCCASliceOutput::outputControlStruct* val)
 937 {
 938         fOutputControl = val;
 939         for (int i = 0;i < fgkNSlices;i++)
 940         {
 941                 fSlaveTrackers[i].SetOutputControl(val);
 942         }
 943 }
 944
 945 #endif