GPU: Move debug starthit sorting to kernel instead of debug dump

ktf · Feb 16, 2024 · 255082a · 255082a
1 parent e90c7d2
commit 255082a
Show file tree

Hide file tree

Showing 6 changed files with 22 additions and 17 deletions.
diff --git a/GPU/GPUTracking/Global/GPUChainTrackingSliceTracker.cxx b/GPU/GPUTracking/Global/GPUChainTrackingSliceTracker.cxx
@@ -247,6 +247,9 @@ int GPUChainTracking::RunTPCTrackingSlices_internal()
       runKernel<GPUTPCStartHitsSorter>(GetGridAuto(useStream), {iSlice});
     }
 #endif
+    if (GetProcessingSettings().comparableDebutOutput) {
+      runKernel<GPUTPCSectorDebugSortKernels, GPUTPCSectorDebugSortKernels::startHits>(GetGrid(1, 1, useStream), {iSlice});
+    }
     DoDebugAndDump(RecoStep::TPCSliceTracking, 32, trk, &GPUTPCTracker::DumpStartHits, *mDebugFile);
 
     if (GetProcessingSettings().memoryAllocationStrategy == GPUMemoryResource::ALLOCATION_INDIVIDUAL) {

diff --git a/GPU/GPUTracking/SliceTracker/GPUTPCSectorDebugSortKernels.cxx b/GPU/GPUTracking/SliceTracker/GPUTPCSectorDebugSortKernels.cxx
@@ -70,3 +70,17 @@ GPUdii() void GPUTPCSectorDebugSortKernels::Thread<GPUTPCSectorDebugSortKernels:
     }
   }
 }
+
+template <>
+GPUdii() void GPUTPCSectorDebugSortKernels::Thread<GPUTPCSectorDebugSortKernels::startHits>(int nBlocks, int nThreads, int iBlock, int iThread, GPUsharedref() GPUSharedMemory& smem, processorType& GPUrestrict() tracker)
+{
+  if (iThread || iBlock) {
+    return;
+  }
+  GPUCommonAlgorithm::sortDeviceDynamic(tracker.TrackletStartHits(), tracker.TrackletStartHits() + *tracker.NStartHits(), [](const GPUTPCHitId& a, const GPUTPCHitId& b) {
+    if (a.RowIndex() != b.RowIndex()) {
+      return (a.RowIndex() < b.RowIndex());
+    }
+    return (a.HitIndex() < b.HitIndex());
+  });
+}
diff --git a/GPU/GPUTracking/SliceTracker/GPUTPCSectorDebugSortKernels.h b/GPU/GPUTracking/SliceTracker/GPUTPCSectorDebugSortKernels.h
@@ -27,7 +27,8 @@ class GPUTPCSectorDebugSortKernels : public GPUKernelTemplate
 {
  public:
   enum K { defaultKernel = 0,
-           hitData = 0 };
+           hitData = 0,
+           startHits = 1 };
   GPUhdi() CONSTEXPR static GPUDataTypes::RecoStep GetRecoStep() { return GPUDataTypes::RecoStep::TPCSliceTracking; }
   typedef GPUTPCTracker processorType;
   GPUhdi() static processorType* Processor(GPUConstantMem& processors) { return processors.tpcTrackers; }

diff --git a/GPU/GPUTracking/SliceTracker/GPUTPCTracker.h b/GPU/GPUTracking/SliceTracker/GPUTPCTracker.h
@@ -230,7 +230,8 @@ class GPUTPCTracker : public GPUProcessor
   GPUhd() GPUglobalref() GPUAtomic(unsigned int) * NStartHits() const { return &mCommonMem->nStartHits; }
 
   GPUhd() GPUglobalref() const GPUTPCHitId& TrackletStartHit(int i) const { return mTrackletStartHits[i]; }
-  GPUhd() GPUglobalref() GPUTPCHitId* TrackletStartHits() const { return mTrackletStartHits; }
+  GPUhd() GPUglobalref() const GPUTPCHitId* TrackletStartHits() const { return mTrackletStartHits; }
+  GPUhd() GPUglobalref() GPUTPCHitId* TrackletStartHits() { return mTrackletStartHits; }
   GPUhd() GPUglobalref() GPUTPCHitId* TrackletTmpStartHits() const { return mTrackletTmpStartHits; }
   MEM_CLASS_PRE2()
   GPUhd() GPUglobalref() const MEM_LG2(GPUTPCTracklet) & Tracklet(int i) const { return mTracklets[i]; }

diff --git a/GPU/GPUTracking/SliceTracker/GPUTPCTrackerDump.cxx b/GPU/GPUTracking/SliceTracker/GPUTPCTrackerDump.cxx
@@ -99,25 +99,10 @@ void GPUTPCTracker::DumpHitWeights(std::ostream& out)
   }
 }
 
-int GPUTPCTracker::StarthitSortComparison(const void* a, const void* b)
-{
-  // qsort helper function to sort start hits
-  const GPUTPCHitId* aa = reinterpret_cast<const GPUTPCHitId*>(a);
-  const GPUTPCHitId* bb = reinterpret_cast<const GPUTPCHitId*>(b);
-
-  if (aa->RowIndex() != bb->RowIndex()) {
-    return (aa->RowIndex() - bb->RowIndex());
-  }
-  return (aa->HitIndex() - bb->HitIndex());
-}
-
 void GPUTPCTracker::DumpStartHits(std::ostream& out)
 {
   // sort start hits and dump to file
   out << "\nStart Hits: (Slice" << mISlice << ") (" << *NStartHits() << ")" << std::endl;
-  if (mRec->GetProcessingSettings().comparableDebutOutput) {
-    qsort(TrackletStartHits(), *NStartHits(), sizeof(GPUTPCHitId), StarthitSortComparison);
-  }
   for (unsigned int i = 0; i < *NStartHits(); i++) {
     out << TrackletStartHit(i).RowIndex() << "-" << TrackletStartHit(i).HitIndex() << std::endl;
   }

diff --git a/GPU/GPUTracking/kernels.cmake b/GPU/GPUTracking/kernels.cmake
@@ -24,6 +24,7 @@ o2_gpu_add_kernel("GPUTPCGlobalTrackingCopyNumbers"              NO_OCL1 single
 o2_gpu_add_kernel("GPUTPCCreateSliceData"                        LB      single)
 o2_gpu_add_kernel("GPUTPCGlobalTracking"                         LB      single)
 o2_gpu_add_kernel("GPUTPCSectorDebugSortKernels, hitData"        NO      single)
+o2_gpu_add_kernel("GPUTPCSectorDebugSortKernels, startHits"      NO      single)
 o2_gpu_add_kernel("GPUTPCCreateOccupancyMap, fill"               LB      simple GPUTPCClusterOccupancyMapBin* map)
 o2_gpu_add_kernel("GPUTPCCreateOccupancyMap, fold"               LB      simple GPUTPCClusterOccupancyMapBin* map)
 o2_gpu_add_kernel("GPUTPCGMMergerTrackFit"                       LB      simple int mode)