GPU: Add possibility to allocate also non-device volatile memory

AliceO2Group · Mar 4, 2024 · 6a28e8d · 6a28e8d
1 parent b4a04fd
commit 6a28e8d
Show file tree

Hide file tree

Showing 3 changed files with 21 additions and 3 deletions.
diff --git a/GPU/GPUTracking/Base/GPUReconstruction.cxx b/GPU/GPUTracking/Base/GPUReconstruction.cxx
@@ -729,6 +729,15 @@ void* GPUReconstruction::AllocateVolatileDeviceMemory(size_t size)
   return retVal;
 }
 
+void* GPUReconstruction::AllocateVolatileMemory(size_t size, bool device)
+{
+  if (device) {
+    return AllocateVolatileDeviceMemory(size);
+  }
+  mVolatileChunks.emplace_back(new char[size + GPUCA_BUFFER_ALIGNMENT]);
+  return GPUProcessor::alignPointer<GPUCA_BUFFER_ALIGNMENT>(mVolatileChunks.back().get());
+}
+
 void GPUReconstruction::ResetRegisteredMemoryPointers(GPUProcessor* proc)
 {
   for (unsigned int i = 0; i < mMemoryResources.size(); i++) {
@@ -796,6 +805,12 @@ void GPUReconstruction::ReturnVolatileDeviceMemory()
   }
 }
 
+void GPUReconstruction::ReturnVolatileMemory()
+{
+  ReturnVolatileDeviceMemory();
+  mVolatileChunks.clear();
+}
+
 void GPUReconstruction::PushNonPersistentMemory(unsigned long tag)
 {
   mNonPersistentMemoryStack.emplace_back(mHostMemoryPoolEnd, mDeviceMemoryPoolEnd, mNonPersistentIndividualAllocations.size(), tag);

diff --git a/GPU/GPUTracking/Base/GPUReconstruction.h b/GPU/GPUTracking/Base/GPUReconstruction.h
@@ -62,6 +62,7 @@ class GPUReconstruction
   std::shared_ptr<LibraryLoader> mMyLib = nullptr;
   std::vector<GPUMemoryResource> mMemoryResources;
   std::vector<std::unique_ptr<char[]>> mUnmanagedChunks;
+  std::vector<std::unique_ptr<char[]>> mVolatileChunks;
   std::vector<std::unique_ptr<GPUChain>> mChains;
 
  public:
@@ -205,10 +206,12 @@ class GPUReconstruction
   void AllocateRegisteredForeignMemory(short res, GPUReconstruction* rec, GPUOutputControl* control = nullptr);
   void* AllocateUnmanagedMemory(size_t size, int type);
   void* AllocateVolatileDeviceMemory(size_t size);
+  void* AllocateVolatileMemory(size_t size, bool device);
   void FreeRegisteredMemory(GPUProcessor* proc, bool freeCustom = false, bool freePermanent = false);
   void FreeRegisteredMemory(short res);
   void ClearAllocatedMemory(bool clearOutputs = true);
   void ReturnVolatileDeviceMemory();
+  void ReturnVolatileMemory();
   void PushNonPersistentMemory(unsigned long tag);
   void PopNonPersistentMemory(RecoStep step, unsigned long tag);
   void BlockStackedMemory(GPUReconstruction* rec);

diff --git a/GPU/GPUTracking/Global/GPUChainTrackingSliceTracker.cxx b/GPU/GPUTracking/Global/GPUChainTrackingSliceTracker.cxx
@@ -152,11 +152,11 @@ int GPUChainTracking::RunTPCTrackingSlices_internal()
     int streamOccMap = mRec->NStreams() - 1;
     runKernel<GPUMemClean16>(GetGridAutoStep(streamOccMap, RecoStep::TPCSliceTracking), krnlRunRangeNone, {}, ptr, GPUTPCClusterOccupancyMapBin::getTotalSize(param()));
     runKernel<GPUTPCCreateOccupancyMap, GPUTPCCreateOccupancyMap::fill>(GetGridBlk(GPUCA_NSLICES * GPUCA_ROW_COUNT, streamOccMap), krnlRunRangeNone, krnlEventNone, ptr);
-    runKernel<GPUTPCCreateOccupancyMap, GPUTPCCreateOccupancyMap::fold>(GetGridBlk(GPUCA_NSLICES * GPUCA_ROW_COUNT, streamOccMap), krnlRunRangeNone, {&mEvents->init}, ptr);
+    runKernel<GPUTPCCreateOccupancyMap, GPUTPCCreateOccupancyMap::fold>(GetGridBlk(GPUCA_NSLICES * GPUCA_ROW_COUNT, streamOccMap), krnlRunRangeNone, krnlEventNone, ptr);
     if (doGPU) {
-      TransferMemoryResourceLinkToHost(RecoStep::TPCSliceTracking, mInputsHost->mResourceOccupancyMap);
+      TransferMemoryResourceLinkToHost(RecoStep::TPCSliceTracking, mInputsHost->mResourceOccupancyMap, streamOccMap, &mEvents->init);
     } else {
-      TransferMemoryResourceLinkToGPU(RecoStep::TPCSliceTracking, mInputsHost->mResourceOccupancyMap);
+      TransferMemoryResourceLinkToGPU(RecoStep::TPCSliceTracking, mInputsHost->mResourceOccupancyMap, streamOccMap, &mEvents->init);
     }
     mRec->UpdateParamOccupancyMap(mInputsHost->mTPCClusterOccupancyMap, mInputsShadow->mTPCClusterOccupancyMap, streamOccMap);
   }