GPU: Cleanup, remove obsolete code

GPU/GPUTracking/Base/GPUReconstructionCPU.h

@@ -112,7 +112,6 @@ class GPUReconstructionCPU : public GPUReconstructionKernels<GPUReconstructionCP
   int32_t ExitDevice() override;
   int32_t GetThread();
 
-  virtual int32_t PrepareTextures() { return 0; }
   virtual int32_t DoStuckProtection(int32_t stream, deviceEvent event) { return 0; }
 
   // Pointers to tracker classes

GPU/GPUTracking/Base/cuda/GPUReconstructionCUDA.cu

@@ -254,12 +254,6 @@ int32_t GPUReconstructionCUDA::InitDevice_Runtime()
       return (1);
     }
 
-#ifdef GPUCA_USE_TEXTURES
-    if (GPUCA_SECTOR_DATA_MEMORY * NSECTORS > (size_t)deviceProp.maxTexture1DLinear) {
-      GPUError("Invalid maximum texture size of device: %ld < %ld\n", (int64_t)deviceProp.maxTexture1DLinear, (int64_t)(GPUCA_SECTOR_DATA_MEMORY * NSECTORS));
-      return (1);
-    }
-#endif
 #ifndef GPUCA_NO_CONSTANT_MEMORY
     if (gGPUConstantMemBufferSize > deviceProp.totalConstMem) {
       GPUError("Insufficient constant memory available on GPU %d < %d!", (int32_t)deviceProp.totalConstMem, (int32_t)gGPUConstantMemBufferSize);
@@ -627,18 +621,6 @@ void GPUReconstructionCUDA::loadKernelModules(bool perKernel)
 }
 
 #ifndef __HIPCC__ // CUDA
-int32_t GPUReconstructionCUDA::PrepareTextures()
-{
-#ifdef GPUCA_USE_TEXTURES
-  cudaChannelFormatDesc channelDescu2 = cudaCreateChannelDesc<cahit2>();
-  size_t offset;
-  GPUChkErr(cudaBindTexture(&offset, &gAliTexRefu2, mProcessorsShadow->tpcTrackers[0].Data().Memory(), &channelDescu2, NSECTORS * GPUCA_SECTOR_DATA_MEMORY));
-  cudaChannelFormatDesc channelDescu = cudaCreateChannelDesc<calink>();
-  GPUChkErr(cudaBindTexture(&offset, &gAliTexRefu, mProcessorsShadow->tpcTrackers[0].Data().Memory(), &channelDescu, NSECTORS * GPUCA_SECTOR_DATA_MEMORY));
-#endif
-  return (0);
-}
-
 void GPUReconstructionCUDA::startGPUProfiling()
 {
   GPUChkErr(cudaProfilerStart());

GPU/GPUTracking/Base/cuda/GPUReconstructionCUDA.h

@@ -79,7 +79,6 @@ class GPUReconstructionCUDA : public GPUReconstructionKernels<GPUReconstructionC
 
 #ifndef __HIPCC__ // CUDA
   bool CanQueryMaxMemory() override { return true; }
-  int32_t PrepareTextures() override;
   void startGPUProfiling() override;
   void endGPUProfiling() override;
 #else // HIP

GPU/GPUTracking/Base/cuda/GPUReconstructionCUDAKernels.cu

@@ -20,11 +20,6 @@
 
 using namespace o2::gpu;
 
-#ifdef GPUCA_USE_TEXTURES
-texture<cahit2, cudaTextureType1D, cudaReadModeElementType> gAliTexRefu2;
-texture<calink, cudaTextureType1D, cudaReadModeElementType> gAliTexRefu;
-#endif
-
 #include "GPUReconstructionIncludesDeviceAll.h"
 
 #if defined(__HIPCC__) && defined(GPUCA_HAS_GLOBAL_SYMBOL_CONSTANT_MEM)

GPU/GPUTracking/Base/opencl/GPUReconstructionOCL.cl

@@ -73,8 +73,6 @@ typedef signed char int8_t;
 #include "GPUConstantMem.h"
 #include "GPUReconstructionIncludesDeviceAll.h"
 
-// if (gpu_mem != pTracker.GPUParametersConst()->gpumem) return; //TODO!
-
 #define GPUCA_KRNL(...) GPUCA_KRNLGPU(__VA_ARGS__)
 #define GPUCA_CONSMEM_PTR GPUglobal() char *gpu_mem, GPUconstant() GPUConstantMem* pConstant,
 #define GPUCA_CONSMEM (*pConstant)

GPU/GPUTracking/Definitions/GPUDef.h

@@ -58,12 +58,6 @@
   #define CA_SHARED_CACHE_REF(target, src, size, reftype, ref) GPUglobalref() const reftype* __restrict__ ref = src
 #endif
 
-#ifdef GPUCA_TEXTURE_FETCH_CONSTRUCTOR
-  #define CA_TEXTURE_FETCH(type, texture, address, entry) tex1Dfetch(texture, ((char*) address - tracker.Data().GPUTextureBase()) / sizeof(type) + entry);
-#else
-  #define CA_TEXTURE_FETCH(type, texture, address, entry) address[entry];
-#endif
-
 #endif //GPUTPCDEF_H
 
 #ifdef GPUCA_CADEBUG

GPU/GPUTracking/Definitions/GPUDefParametersDefaults.h

@@ -267,7 +267,6 @@
     #define GPUCA_PAR_COMP_GATHER_MODE 3
     #define GPUCA_PAR_DEDX_STORAGE_TYPE uint16_t
     #define GPUCA_PAR_MERGER_INTERPOLATION_ERROR_TYPE half
-    // #define GPUCA_USE_TEXTURES
   #elif defined(GPUCA_GPUTYPE_OPENCL)
   #else
     #error GPU TYPE NOT SET

GPU/GPUTracking/Definitions/GPUDefParametersWrapper.h

@@ -43,11 +43,6 @@ namespace o2::gpu
 
 // #define GPUCA_KERNEL_DEBUGGER_OUTPUT
 
-// Derived parameters
-#ifdef GPUCA_USE_TEXTURES
-  #define GPUCA_TEXTURE_FETCH_CONSTRUCTOR                              // Fetch data through texture cache
-#endif
-
 } // namespace o2::gpu
 
 // clang-format on

GPU/GPUTracking/Global/GPUChain.h

@@ -216,7 +216,6 @@ class GPUChain
 
   inline GPUChain* GetNextChainInQueue() { return mRec->GetNextChainInQueue(); }
 
-  virtual int32_t PrepareTextures() { return 0; }
   virtual int32_t DoStuckProtection(int32_t stream, deviceEvent event) { return 0; }
 
   template <class T, class S, typename... Args>

GPU/GPUTracking/Global/GPUChainTrackingSectorTracker.cxx

@@ -103,17 +103,6 @@ int32_t GPUChainTracking::RunTPCTrackingSectors_internal()
   int32_t streamInitAndOccMap = mRec->NStreams() - 1;
 
   if (doGPU) {
-    for (uint32_t iSector = 0; iSector < NSECTORS; iSector++) {
-      processorsShadow()->tpcTrackers[iSector].GPUParametersConst()->gpumem = (char*)mRec->DeviceMemoryBase();
-      // Initialize Startup Constants
-      processors()->tpcTrackers[iSector].GPUParameters()->nextStartHit = (((getKernelProperties<GPUTPCTrackletConstructor>().minBlocks * BlockCount()) + NSECTORS - 1 - iSector) / NSECTORS) * getKernelProperties<GPUTPCTrackletConstructor>().nThreads;
-      processorsShadow()->tpcTrackers[iSector].SetGPUTextureBase(mRec->DeviceMemoryBase());
-    }
-
-    if (PrepareTextures()) {
-      return (2);
-    }
-
     // Copy Tracker Object to GPU Memory
     if (GetProcessingSettings().debugLevel >= 3) {
       GPUInfo("Copying Tracker objects to GPU");

GPU/GPUTracking/Merger/GPUTPCGMTrackParam.cxx

@@ -580,10 +580,8 @@ GPUd() float GPUTPCGMTrackParam::AttachClusters(const GPUTPCGMMerger* GPUrestric
   }
   const GPUTPCTracker& GPUrestrict() tracker = *(Merger -> GetConstantMem()->tpcTrackers + sector);
   const GPUTPCRow& GPUrestrict() row = tracker.Row(iRow);
-#ifndef GPUCA_TEXTURE_FETCH_CONSTRUCTOR
   GPUglobalref() const cahit2* hits = tracker.HitData(row);
   GPUglobalref() const calink* firsthit = tracker.FirstHitInBin(row);
-#endif //! GPUCA_TEXTURE_FETCH_CONSTRUCTOR
   if (row.NHits() == 0) {
     return -1e6f;
   }
@@ -626,8 +624,8 @@ GPUd() float GPUTPCGMTrackParam::AttachClusters(const GPUTPCGMMerger* GPUrestric
   }
   for (int32_t k = 0; k <= nz; k++) {
     const int32_t mybin = bin + k * nBinsY;
-    const uint32_t hitFst = CA_TEXTURE_FETCH(calink, gAliTexRefu, firsthit, mybin);
-    const uint32_t hitLst = CA_TEXTURE_FETCH(calink, gAliTexRefu, firsthit, mybin + ny + 1);
+    const uint32_t hitFst = firsthit[mybin];
+    const uint32_t hitLst = firsthit[mybin + ny + 1];
     for (uint32_t ih = hitFst; ih < hitLst; ih++) {
       int32_t id = idOffset + ids[ih];
       GPUAtomic(uint32_t)* const weight = weights + id;
@@ -636,7 +634,7 @@ GPUd() float GPUTPCGMTrackParam::AttachClusters(const GPUTPCGMMerger* GPUrestric
           continue;
         }
       }
-      const cahit2 hh = CA_TEXTURE_FETCH(cahit2, gAliTexRefu2, hits, ih);
+      const cahit2 hh = hits[ih];
       const float y = y0 + hh.x * stepY;
       const float z = z0 + hh.y * stepZ;
       const float dy = y - uncorrectedY;

GPU/GPUTracking/SectorTracker/GPUTPCTracker.h

@@ -59,24 +59,14 @@ class GPUTPCTracker : public GPUProcessor
   void DumpTrackletHits(std::ostream& out);         // Same for Track Hits
 #endif
 
-  struct StructGPUParameters {
-    GPUAtomic(uint32_t) nextStartHit; // Next Tracklet to process
-  };
-
-  struct StructGPUParametersConst {
-    GPUglobalref() char* gpumem; // Base pointer to GPU memory (Needed for OpenCL for verification)
-  };
-
   struct commonMemoryStruct {
-    commonMemoryStruct() : nStartHits(0), nTracklets(0), nRowHits(0), nTracks(0), nLocalTracks(0), nTrackHits(0), nLocalTrackHits(0), gpuParameters() {}
-    GPUAtomic(uint32_t) nStartHits;    // number of start hits
-    GPUAtomic(uint32_t) nTracklets;    // number of tracklets
-    GPUAtomic(uint32_t) nRowHits;      // number of tracklet hits
-    GPUAtomic(uint32_t) nTracks;       // number of reconstructed tracks
-    int32_t nLocalTracks;              // number of reconstructed tracks before extrapolation tracking
-    GPUAtomic(uint32_t) nTrackHits;    // number of track hits
-    int32_t nLocalTrackHits;           // see above
-    StructGPUParameters gpuParameters; // GPU parameters
+    GPUAtomic(uint32_t) nStartHits = 0; // number of start hits
+    GPUAtomic(uint32_t) nTracklets = 0; // number of tracklets
+    GPUAtomic(uint32_t) nRowHits = 0;   // number of tracklet hits
+    GPUAtomic(uint32_t) nTracks = 0;    // number of reconstructed tracks
+    int32_t nLocalTracks = 0;           // number of reconstructed tracks before extrapolation tracking
+    GPUAtomic(uint32_t) nTrackHits = 0; // number of track hits
+    int32_t nLocalTrackHits = 0;        // see above
   };
 
   GPUhdi() GPUglobalref() const GPUTPCClusterData* ClusterData() const
@@ -205,13 +195,6 @@ class GPUTPCTracker : public GPUProcessor
 
   GPUhd() GPUglobalref() GPUTPCRow* TrackingDataRows() const { return (mData.Rows()); }
   GPUhd() GPUglobalref() int32_t* RowStartHitCountOffset() const { return (mRowStartHitCountOffset); }
-  GPUhd() GPUglobalref() StructGPUParameters* GPUParameters() const { return (&mCommonMem->gpuParameters); }
-  GPUhd() StructGPUParametersConst* GPUParametersConst()
-  {
-    return (&mGPUParametersConst);
-  }
-  GPUhd() const StructGPUParametersConst* GetGPUParametersConst() const { return (&mGPUParametersConst); }
-  GPUhd() void SetGPUTextureBase(GPUglobalref() const void* val) { mData.SetGPUTextureBase(val); }
 
   struct trackSortData {
     int32_t fTtrack; // Track ID
@@ -253,8 +236,6 @@ class GPUTPCTracker : public GPUProcessor
   GPUglobalref() GPUTPCHitId* mTrackletTmpStartHits = nullptr; // Unsorted start hits
   GPUglobalref() char* mGPUTrackletTemp = nullptr;             // Temp Memory for GPU Tracklet Constructor
 
-  StructGPUParametersConst mGPUParametersConst; // Parameters for GPU if this is a GPU tracker
-
   // event
   GPUglobalref() commonMemoryStruct* mCommonMem = nullptr;  // common event memory
   GPUglobalref() GPUTPCHitId* mTrackletStartHits = nullptr; // start hits for the tracklets

GPU/GPUTracking/SectorTracker/GPUTPCTrackingData.h

@@ -29,7 +29,7 @@ class GPUTPCHit;
 class GPUTPCTrackingData
 {
  public:
-  GPUTPCTrackingData() : mNumberOfHits(0), mNumberOfHitsPlusAlign(0), mClusterIdOffset(0), mGPUTextureBase(nullptr), mRows(nullptr), mLinkUpData(nullptr), mLinkDownData(nullptr), mClusterData(nullptr) {}
+  GPUTPCTrackingData() : mNumberOfHits(0), mNumberOfHitsPlusAlign(0), mClusterIdOffset(0), mRows(nullptr), mLinkUpData(nullptr), mLinkDownData(nullptr), mClusterData(nullptr) {}
 
 #ifndef GPUCA_GPUCODE_DEVICE
   ~GPUTPCTrackingData() = default;
@@ -112,10 +112,6 @@ class GPUTPCTrackingData
 
   GPUhdi() GPUglobalref() GPUAtomic(uint32_t) * HitWeights() { return (mHitWeights); }
 
-  GPUhdi() void SetGPUTextureBase(GPUglobalref() const void* val) { mGPUTextureBase = val; }
-  GPUhdi() char* GPUTextureBase() const { return ((char*)mGPUTextureBase); }
-  GPUhdi() char* GPUTextureBaseConst() const { return ((char*)mGPUTextureBase); }
-
   GPUhdi() GPUglobalref() const GPUTPCClusterData* ClusterData() const { return mClusterData; }
 
  private:
@@ -135,8 +131,6 @@ class GPUTPCTrackingData
   int32_t mNumberOfHitsPlusAlign;
   int32_t mClusterIdOffset;
 
-  GPUglobalref() const void* mGPUTextureBase; // pointer to start of GPU texture
-
   GPUglobalref() GPUTPCRow* mRows; // The row objects needed for most accessor functions
 
   GPUglobalref() calink* mLinkUpData;        // hit index in the row above which is linked to the given (global) hit index

GPU/GPUTracking/SectorTracker/GPUTPCTrackletConstructor.cxx

@@ -123,10 +123,10 @@ GPUdic(2, 1) void GPUTPCTrackletConstructor::UpdateTracklet(int32_t /*nBlocks*/,
         break; // SG!!! - jump over the row
       }
 
-      cahit2 hh = CA_TEXTURE_FETCH(cahit22, gAliTexRefu2, tracker.HitData(row), r.mCurrIH);
+      cahit2 hh = tracker.HitData(row)[r.mCurrIH];
 
       int32_t seedIH = r.mCurrIH;
-      r.mCurrIH = CA_TEXTURE_FETCH(calink, gAliTexRefs, tracker.HitLinkUpData(row), r.mCurrIH);
+      r.mCurrIH = tracker.HitLinkUpData(row)[r.mCurrIH];
 
       float x = row.X();
       float y = y0 + hh.x * stepY;
@@ -282,10 +282,8 @@ GPUdic(2, 1) void GPUTPCTrackletConstructor::UpdateTracklet(int32_t /*nBlocks*/,
           break;
         }
 
-#ifndef GPUCA_TEXTURE_FETCH_CONSTRUCTOR
         GPUglobalref() const cahit2* hits = tracker.HitData(row);
         GPUglobalref() const calink* firsthit = tracker.FirstHitInBin(row);
-#endif //! GPUCA_TEXTURE_FETCH_CONSTRUCTOR
         tracker.GetConstantMem()->calibObjects.fastTransformHelper->InverseTransformYZtoNominalYZ(tracker.ISector(), iRow, yUncorrected, zUncorrected, yUncorrected, zUncorrected);
 
         if (tracker.Param().rec.tpc.rejectEdgeClustersInSeeding && tracker.Param().rejectEdgeClusterByY(yUncorrected, iRow, CAMath::Sqrt(tParam.Err2Y()))) {
@@ -318,14 +316,14 @@ GPUdic(2, 1) void GPUTPCTrackletConstructor::UpdateTracklet(int32_t /*nBlocks*/,
 #endif
             int32_t nBinsY = row.Grid().Ny();
             int32_t mybin = bin + k * nBinsY;
-            uint32_t hitFst = CA_TEXTURE_FETCH(calink, gAliTexRefu, firsthit, mybin);
-            uint32_t hitLst = CA_TEXTURE_FETCH(calink, gAliTexRefu, firsthit, mybin + ny + 1);
+            uint32_t hitFst = firsthit[mybin];
+            uint32_t hitLst = firsthit[mybin + ny + 1];
 #ifdef __HIPCC__ // Todo: fixme!
             for (uint32_t ih = hitFst - 1; ++ih < hitLst; /*ih++*/) {
 #else
             for (uint32_t ih = hitFst; ih < hitLst; ih++) {
 #endif
-              cahit2 hh = CA_TEXTURE_FETCH(cahit2, gAliTexRefu2, hits, ih);
+              cahit2 hh = hits[ih];
               float y = y0 + hh.x * stepY;
               float z = z0 + hh.y * stepZ;
               float dy = y - yUncorrected;
@@ -353,7 +351,7 @@ GPUdic(2, 1) void GPUTPCTrackletConstructor::UpdateTracklet(int32_t /*nBlocks*/,
           }
         }
 
-        cahit2 hh = CA_TEXTURE_FETCH(cahit2, gAliTexRefu2, hits, best);
+        cahit2 hh = hits[best];
         float y = y0 + hh.x * stepY + tParam.GetY() - yUncorrected;
         float z = z0 + hh.y * stepZ + tParam.GetZ() - zUncorrected;
 
@@ -390,8 +388,8 @@ GPUdic(2, 1) void GPUTPCTrackletConstructor::UpdateTracklet(int32_t /*nBlocks*/,
     const GPUglobalref() GPUTPCRow& GPUrestrict() row2 = tracker.Row(r.mLastRow);
     GPUglobalref() const cahit2* hits1 = tracker.HitData(row1);
     GPUglobalref() const cahit2* hits2 = tracker.HitData(row2);
-    const cahit2 hh1 = CA_TEXTURE_FETCH(cahit2, gAliTexRefu2, hits1, rowHits[r.mFirstRow]);
-    const cahit2 hh2 = CA_TEXTURE_FETCH(cahit2, gAliTexRefu2, hits2, rowHits[r.mLastRow]);
+    const cahit2 hh1 = hits1[rowHits[r.mFirstRow]];
+    const cahit2 hh2 = hits2[rowHits[r.mLastRow]];
     const float z1 = row1.Grid().ZMin() + hh1.y * row1.HstepZ();
     const float z2 = row2.Grid().ZMin() + hh2.y * row2.HstepZ();
     float oldOffset = tParam.ZOffset();

GPU/GPUTracking/SectorTracker/GPUTPCTrackletConstructor.h

@@ -63,9 +63,6 @@ class GPUTPCTrackletConstructor : public GPUKernelTemplate
 
   struct GPUSharedMemory {
     CA_SHARED_STORAGE(GPUTPCRow mRows[GPUCA_ROW_COUNT]); // rows
-    int32_t mNextStartHitFirst;                          // First start hit to be processed by CUDA block during next iteration
-    int32_t mNextStartHitCount;                          // Number of start hits to be processed by CUDA block during next iteration
-    int32_t mNextStartHitFirstRun;                       // First run for dynamic scheduler?
     int32_t mNStartHits;                                 // Total number of start hits
 
 #ifdef GPUCA_TRACKLET_CONSTRUCTOR_DO_PROFILE