Works except for two partial template specializations

Author: OliverRietmann

GPU/GPUTracking/DataTypes/GPUDataTypes.h

@@ -24,6 +24,7 @@
 #include <cstddef>
 #endif
 #include "GPUTRDDef.h"
+#include "GPUTPCDef.h"
 
 struct AliHLTTPCClusterMCLabel;
 struct AliHLTTPCRawCluster;
@@ -97,7 +98,8 @@ struct TPCZSLinkMapping;
 
 #include "utils/bitfield.h"
 
-class GPUTPCTrack;
+template <template <class> class F>
+struct GPUTPCTrackSkeleton;
 class GPUTPCHitId;
 class GPUTPCGMMergedTrack;
 struct GPUTPCGMMergedTrackHit;
@@ -224,7 +226,7 @@ struct GPUTrackingInOutPointers {
   const AliHLTTPCRawCluster* rawClusters[NSECTORS] = {nullptr};
   uint32_t nRawClusters[NSECTORS] = {0};
   const o2::tpc::ClusterNativeAccess* clustersNative = nullptr;
-  const GPUTPCTrack* sectorTracks[NSECTORS] = {nullptr};
+  const GPUTPCTrackSkeleton<wrapper::value>* sectorTracks[NSECTORS] = {nullptr};  // GPUTPCTrack
   uint32_t nSectorTracks[NSECTORS] = {0};
   const GPUTPCHitId* sectorClusters[NSECTORS] = {nullptr};
   uint32_t nSectorClusters[NSECTORS] = {0};

GPU/GPUTracking/Global/GPUChainTracking.h

@@ -107,7 +107,7 @@ class GPUChainTracking : public GPUChain
     std::unique_ptr<AliHLTTPCRawCluster[]> rawClusters[NSECTORS];
     std::unique_ptr<o2::tpc::ClusterNative[]> clustersNative;
     std::unique_ptr<o2::tpc::ClusterNativeAccess> clusterNativeAccess;
-    std::unique_ptr<GPUTPCTrack[]> sectorTracks[NSECTORS];
+    std::unique_ptr<GPUTPCTrackSkeleton<wrapper::value>[]> sectorTracks[NSECTORS];
     std::unique_ptr<GPUTPCHitId[]> sectorClusters[NSECTORS];
     std::unique_ptr<AliHLTTPCClusterMCLabel[]> mcLabelsTPC;
     std::unique_ptr<GPUTPCMCInfo[]> mcInfosTPC;

GPU/GPUTracking/SectorTracker/GPUTPCBaseTrackParam.h

@@ -16,7 +16,7 @@
 #define GPUTPCBASETRACKPARAM_H
 
 #include "GPUTPCDef.h"
-#include "wrapper.h"
+// #include "wrapper.h"
 
 namespace o2::gpu
 {
@@ -30,6 +30,12 @@ namespace o2::gpu
  */
 template <template <class> class F>
 struct GPUTPCBaseTrackParamSkeleton {
+  template <template <class> class F_new>
+  operator GPUTPCBaseTrackParamSkeleton<F_new>() const { return {mX, mC, mZOffset, mP}; }
+
+  template <template <class> class F_new>
+  operator GPUTPCBaseTrackParamSkeleton<F_new>() { return {mX, mC, mZOffset, mP}; }
+
   GPUd() float X() const { return mX; }
   GPUd() float Y() const { return mP[0]; }
   GPUd() float Z() const { return mP[1]; }

GPU/GPUTracking/SectorTracker/GPUTPCDef.h

@@ -31,6 +31,14 @@ typedef uint32_t calink;
 typedef uint32_t cahit;
 #endif
 struct cahit2 { cahit x, y; };
+
+namespace wrapper {
+    template <class T> using value = T;
+    template <class T> using reference = T&;
+    template <class T> using const_reference = const T&;
+    template <class T> using pointer = T*;
+} // namespace o2::gpu::wrapper
+
 } // namespace o2::GPU
 
 #endif //GPUDTPCEF_H

GPU/GPUTracking/SectorTracker/GPUTPCTrack.h

@@ -17,6 +17,7 @@
 
 #include "GPUTPCBaseTrackParam.h"
 #include "GPUTPCDef.h"
+// #include "wrapper.h"
 
 namespace o2::gpu
 {
@@ -27,14 +28,13 @@ namespace o2::gpu
  * The class is dedicated for internal use by the GPUTPCTracker algorithm.
  * The track parameters at both ends are stored separately in the GPUTPCEndPoint class
  */
-class GPUTPCTrack
+template <template <class> class F>
+class GPUTPCTrackSkeleton
 {
  public:
 #if !defined(GPUCA_GPUCODE)
-  GPUTPCTrack() : mFirstHitID(0), mNHits(0), mLocalTrackId(-1), mParam()
-  {
-  }
-  ~GPUTPCTrack() = default;
+  GPUTPCTrackSkeleton();
+  ~GPUTPCTrackSkeleton() = default;
 #endif //! GPUCA_GPUCODE
 
   GPUhd() int32_t NHits() const { return mNHits; }
@@ -45,8 +45,30 @@ class GPUTPCTrack
   GPUhd() void SetNHits(int32_t v) { mNHits = v; }
   GPUhd() void SetLocalTrackId(int32_t v) { mLocalTrackId = v; }
   GPUhd() void SetFirstHitID(int32_t v) { mFirstHitID = v; }
-
-  GPUhd() void SetParam(const GPUTPCBaseTrackParam& v) { mParam = v; }
+  GPUhd() void SetParam(GPUTPCBaseTrackParamSkeleton<wrapper::const_reference> v) {
+    mParam.mX = v.mX;
+    mParam.mC[0] = v.mC[0];
+    mParam.mC[1] = v.mC[1];
+    mParam.mC[2] = v.mC[2];
+    mParam.mC[3] = v.mC[3];
+    mParam.mC[4] = v.mC[4];
+    mParam.mC[5] = v.mC[5];
+    mParam.mC[6] = v.mC[6];
+    mParam.mC[7] = v.mC[7];
+    mParam.mC[8] = v.mC[8];
+    mParam.mC[9] = v.mC[9];
+    mParam.mC[10] = v.mC[10];
+    mParam.mC[11] = v.mC[11];
+    mParam.mC[12] = v.mC[12];
+    mParam.mC[13] = v.mC[13];
+    mParam.mC[14] = v.mC[14];
+    mParam.mZOffset = v.mZOffset;
+    mParam.mP[0] = v.mP[0];
+    mParam.mP[1] = v.mP[1];
+    mParam.mP[2] = v.mP[2];
+    mParam.mP[3] = v.mP[3];
+    mParam.mP[4] = v.mP[4];
+  }
 
  private:
   int32_t mFirstHitID;         // index of the first track cell in the track->cell pointer array
@@ -56,6 +78,16 @@ class GPUTPCTrack
 
  private:
 };
+
+#if !defined(GPUCA_GPUCODE)
+template <>
+GPUTPCTrackSkeleton<wrapper::value>::GPUTPCTrackSkeleton() : mFirstHitID(0), mNHits(0), mLocalTrackId(-1), mParam()
+{
+}
+#endif //! GPUCA_GPUCODE
+
+using GPUTPCTrack = GPUTPCTrackSkeleton<wrapper::value>;
+
 } // namespace o2::gpu
 
 #endif // GPUTPCTRACK_H

GPU/GPUTracking/SectorTracker/GPUTPCTrackParam.cxx

@@ -15,7 +15,8 @@
 #include "GPUTPCTrackLinearisation.h"
 #include "GPUTPCTrackParam.h"
 #include "GPUTPCGeometry.h"
-#include "wrapper.h"
+#include "GPUTPCDef.h"
+// #include "wrapper.h"
 
 using namespace o2::gpu;
 

GPU/GPUTracking/SectorTracker/GPUTPCTrackParam.h

@@ -18,7 +18,7 @@
 #include "GPUTPCBaseTrackParam.h"
 #include "GPUTPCDef.h"
 #include "GPUCommonMath.h"
-#include "wrapper.h"
+// #include "wrapper.h"
 
 namespace o2::gpu
 {

GPU/GPUTracking/SectorTracker/GPUTPCTracker.h

@@ -31,7 +31,8 @@ namespace o2::gpu
 {
 struct GPUTPCClusterData;
 struct GPUParam;
-class GPUTPCTrack;
+template <template <class> class F>
+struct GPUTPCTrackSkeleton;
 class GPUTPCRow;
 
 class GPUTPCTracker : public GPUProcessor
@@ -201,7 +202,7 @@ class GPUTPCTracker : public GPUProcessor
   GPUhd() GPUglobalref() calink* TrackletRowHits() const { return mTrackletRowHits; }
 
   GPUhd() GPUglobalref() GPUAtomic(uint32_t) * NTracks() const { return &mCommonMem->nTracks; }
-  GPUhd() GPUglobalref() GPUTPCTrack* Tracks() const { return mTracks; }
+  GPUhd() GPUglobalref() GPUTPCTrackSkeleton<wrapper::value>* Tracks() const { return mTracks; }
   GPUhd() GPUglobalref() GPUAtomic(uint32_t) * NTrackHits() const { return &mCommonMem->nTrackHits; }
   GPUhd() GPUglobalref() GPUTPCHitId* TrackHits() const { return mTrackHits; }
 
@@ -253,7 +254,7 @@ class GPUTPCTracker : public GPUProcessor
   GPUglobalref() GPUTPCHitId* mTrackletStartHits = nullptr; // start hits for the tracklets
   GPUglobalref() GPUTPCTracklet_pointer mTracklets;// tracklets
   GPUglobalref() calink* mTrackletRowHits = nullptr;        // Hits for each Tracklet in each row
-  GPUglobalref() GPUTPCTrack* mTracks = nullptr;            // reconstructed tracks
+  GPUglobalref() GPUTPCTrackSkeleton<wrapper::value>* mTracks = nullptr;            // reconstructed tracks
   GPUglobalref() GPUTPCHitId* mTrackHits = nullptr;         // array of track hit numbers
 
   static int32_t StarthitSortComparison(const void* a, const void* b);

GPU/GPUTracking/SectorTracker/GPUTPCTracklet.h

@@ -17,7 +17,7 @@
 
 #include "GPUTPCBaseTrackParam.h"
 #include "GPUTPCDef.h"
-#include "wrapper.h"
+// #include "wrapper.h"
 
 namespace o2::gpu
 {
@@ -39,12 +39,35 @@ class GPUTPCTrackletSkeleton
   GPUhd() int32_t LastRow() const { return mLastRow; }
   GPUhd() int32_t HitWeight() const { return mHitWeight; }
   GPUhd() uint32_t FirstHit() const { return mFirstHit; }
-  GPUhd() const GPUTPCBaseTrackParam& Param() const { return mParam; }
+  GPUhd() GPUTPCBaseTrackParamSkeleton<wrapper::const_reference> Param() const { return {mParam.mX, mParam.mC, mParam.mZOffset, mParam.mP}; }
 
   GPUhd() void SetFirstRow(int32_t v) { mFirstRow = v; }
   GPUhd() void SetLastRow(int32_t v) { mLastRow = v; }
   GPUhd() void SetFirstHit(uint32_t v) { mFirstHit = v; }
-  GPUhd() void SetParam(const GPUTPCBaseTrackParam& v) { mParam = reinterpret_cast<const GPUTPCBaseTrackParam&>(v); }
+  GPUhd() void SetParam(GPUTPCBaseTrackParamSkeleton<wrapper::const_reference> v) {
+    mParam.mX = v.mX;
+    mParam.mC[0] = v.mC[0];
+    mParam.mC[1] = v.mC[1];
+    mParam.mC[2] = v.mC[2];
+    mParam.mC[3] = v.mC[3];
+    mParam.mC[4] = v.mC[4];
+    mParam.mC[5] = v.mC[5];
+    mParam.mC[6] = v.mC[6];
+    mParam.mC[7] = v.mC[7];
+    mParam.mC[8] = v.mC[8];
+    mParam.mC[9] = v.mC[9];
+    mParam.mC[10] = v.mC[10];
+    mParam.mC[11] = v.mC[11];
+    mParam.mC[12] = v.mC[12];
+    mParam.mC[13] = v.mC[13];
+    mParam.mC[14] = v.mC[14];
+    mParam.mZOffset = v.mZOffset;
+    mParam.mP[0] = v.mP[0];
+    mParam.mP[1] = v.mP[1];
+    mParam.mP[2] = v.mP[2];
+    mParam.mP[3] = v.mP[3];
+    mParam.mP[4] = v.mP[4];
+  }
   GPUhd() void SetHitWeight(const int32_t w) { mHitWeight = w; }
 
 // private:
@@ -55,6 +78,9 @@ class GPUTPCTrackletSkeleton
   F<uint32_t> mFirstHit;          // first hit in row hit array
 };
 
+template<>
+GPUhd() GPUTPCBaseTrackParamSkeleton<wrapper::const_reference> GPUTPCTrackletSkeleton<wrapper::pointer>::Param() const { return {mParam->mX, mParam->mC, mParam->mZOffset, mParam->mP}; }
+
 using GPUTPCTracklet = GPUTPCTrackletSkeleton<wrapper::value>;
 using GPUTPCTracklet_reference = GPUTPCTrackletSkeleton<wrapper::reference>;
 using GPUTPCTracklet_const_reference = GPUTPCTrackletSkeleton<wrapper::const_reference>;

GPU/GPUTracking/Standalone/cmake/config.cmake

@@ -28,7 +28,7 @@ set(GPUCA_CONFIG_GL3W 0)
 set(GPUCA_CONFIG_O2 1)
 set(GPUCA_BUILD_DEBUG 0)
 set(GPUCA_BUILD_DEBUG_SANITIZE 0)
-set(GPUCA_DETERMINISTIC_MODE 0 CACHE BOOL "GPUCA_DETERMINISTIC_MODE")             # OFF / NO_FAST_MATH / OPTO2 / GPU / WHOLEO2
+set(GPUCA_DETERMINISTIC_MODE 0 CACHE BOOL "GPUCA_DETERMINISTIC_MODE") # OFF / NO_FAST_MATH / OPTO2 / GPU / WHOLEO2
 #set(GPUCA_CUDA_GCCBIN c++-14)
 #set(GPUCA_OPENCL_CLANGBIN clang-20)
 set(HIP_AMDGPUTARGET "default" CACHE STRING "HIP_AMDGPUTARGET")      # "gfx906;gfx908;gfx90a"