GPU: Add option to register all temporary input memory used by the standalone benchmark

Author: davidrohr

GPU/GPUTracking/Base/GPUReconstruction.h

@@ -342,6 +342,9 @@ inline void GPUReconstruction::AllocateIOMemoryHelper(unsigned int n, const T*&
   }
   u.reset(new T[n]);
   ptr = u.get();
+  if (mDeviceProcessingSettings.registerStandaloneInputMemory) {
+    registerMemoryForGPU(u.get(), n * sizeof(T));
+  }
 }
 
 template <class T, typename... Args>

GPU/GPUTracking/Base/GPUReconstructionConvert.cxx

@@ -180,7 +180,7 @@ void GPUReconstructionConvert::ZSstreamOut(unsigned short* bufIn, unsigned int&
   lenIn = 0;
 }
 
-void GPUReconstructionConvert::RunZSEncoder(const GPUTrackingInOutDigits* in, const GPUTrackingInOutZS*& out, const GPUParam& param, bool zs12bit)
+void GPUReconstructionConvert::RunZSEncoder(const GPUTrackingInOutDigits* in, GPUTrackingInOutZS*& out, const GPUParam& param, bool zs12bit)
 {
 #ifdef GPUCA_TPC_GEOMETRY_O2
   static std::vector<std::array<long long int, TPCZSHDR::TPC_ZS_PAGE_SIZE / sizeof(long long int)>> buffer[NSLICES][GPUTrackingInOutZS::NENDPOINTS];

GPU/GPUTracking/Base/GPUReconstructionConvert.h

@@ -48,7 +48,7 @@ class GPUReconstructionConvert
   constexpr static unsigned int NSLICES = GPUCA_NSLICES;
   static void ConvertNativeToClusterData(o2::tpc::ClusterNativeAccess* native, std::unique_ptr<GPUTPCClusterData[]>* clusters, unsigned int* nClusters, const TPCFastTransform* transform, int continuousMaxTimeBin = 0);
   static void ConvertRun2RawToNative(o2::tpc::ClusterNativeAccess& native, std::unique_ptr<o2::tpc::ClusterNative[]>& nativeBuffer, const AliHLTTPCRawCluster** rawClusters, unsigned int* nRawClusters);
-  static void RunZSEncoder(const GPUTrackingInOutDigits* in, const GPUTrackingInOutZS*& out, const GPUParam& param, bool zs12bit);
+  static void RunZSEncoder(const GPUTrackingInOutDigits* in, GPUTrackingInOutZS*& out, const GPUParam& param, bool zs12bit);
   static void RunZSFilter(std::unique_ptr<deprecated::PackedDigit[]>* buffers, const deprecated::PackedDigit* const* ptrs, size_t* nsb, const size_t* ns, const GPUParam& param, bool zs12bit);
   static int GetMaxTimeBin(const o2::tpc::ClusterNativeAccess& native);
   static int GetMaxTimeBin(const GPUTrackingInOutDigits& digits);

GPU/GPUTracking/Base/GPUSettings.cxx

@@ -87,4 +87,5 @@ void GPUSettingsDeviceProcessing::SetDefaults()
   trackletSelectorInPipeline = false;
   forceMemoryPoolSize = 0;
   nTPCClustererLanes = 3;
+  registerStandaloneInputMemory = false;
 }

GPU/GPUTracking/Base/GPUSettings.h

@@ -143,6 +143,7 @@ struct GPUSettingsDeviceProcessing {
   size_t forceMemoryPoolSize;         // Override size of memory pool to be allocated on GPU / Host (set =1 to force allocating all device memory, if supported)
   int nTPCClustererLanes;             // Number of TPC clusterers that can run in parallel
   bool deviceTimers;                  // Use device timers instead of host-based timers
+  bool registerStandaloneInputMemory; // Automatically register memory for the GPU which is used as input for the standalone benchmark
 };
 } // namespace gpu
 } // namespace GPUCA_NAMESPACE

GPU/GPUTracking/Global/GPUChainTracking.cxx

@@ -691,6 +691,9 @@ void GPUChainTracking::ConvertNativeToClusterDataLegacy()
   GPUReconstructionConvert::ConvertNativeToClusterData(mClusterNativeAccess.get(), mIOMem.clusterData, mIOPtrs.nClusterData, processors()->calibObjects.fastTransform, param().continuousMaxTimeBin);
   for (unsigned int i = 0; i < NSLICES; i++) {
     mIOPtrs.clusterData[i] = mIOMem.clusterData[i].get();
+    if (GetDeviceProcessingSettings().registerStandaloneInputMemory) {
+      mRec->registerMemoryForGPU(mIOMem.clusterData[i].get(), mIOPtrs.nClusterData[i] * sizeof(*mIOPtrs.clusterData[i]));
+    }
   }
   mIOPtrs.clustersNative = nullptr;
   mIOMem.clustersNative.reset(nullptr);
@@ -709,11 +712,25 @@ void GPUChainTracking::ConvertRun2RawToNative()
     mIOMem.clusterData[i].reset(nullptr);
   }
   mIOPtrs.clustersNative = mClusterNativeAccess.get();
+  if (GetDeviceProcessingSettings().registerStandaloneInputMemory) {
+    mRec->registerMemoryForGPU(mIOMem.clustersNative.get(), mClusterNativeAccess->nClustersTotal * sizeof(*mClusterNativeAccess->clustersLinear));
+  }
 }
 
 void GPUChainTracking::ConvertZSEncoder(bool zs12bit)
 {
-  GPUReconstructionConvert::RunZSEncoder(mIOPtrs.tpcPackedDigits, mIOPtrs.tpcZS, param(), zs12bit);
+  GPUTrackingInOutZS* tmp;
+  GPUReconstructionConvert::RunZSEncoder(mIOPtrs.tpcPackedDigits, tmp, param(), zs12bit);
+  mIOPtrs.tpcZS = tmp;
+  if (GetDeviceProcessingSettings().registerStandaloneInputMemory) {
+    for (unsigned int i = 0; i < NSLICES; i++) {
+      for (unsigned int j = 0; j < GPUTrackingInOutZS::NENDPOINTS; j++) {
+        for (unsigned int k = 0; k < tmp->slice[i].count[j]; k++) {
+          mRec->registerMemoryForGPU(tmp->slice[i].zsPtr[j][k], tmp->slice[i].nZSPtr[j][k] * TPCZSHDR::TPC_ZS_PAGE_SIZE);
+        }
+      }
+    }
+  }
 }
 
 void GPUChainTracking::ConvertZSFilter(bool zs12bit)

GPU/GPUTracking/Standalone/qconfigoptions.h

@@ -138,6 +138,7 @@ AddOption(dzdr, float, 2.5f, "DzDr", 0, "Use dZ/dR search window instead of vert
 AddOption(cont, bool, false, "continuous", 0, "Process continuous timeframe data")
 AddOption(forceMemorySize, unsigned long long int, 1, "memSize", 0, "Force size of allocated GPU / page locked host memory", min(0ull))
 AddOption(outputcontrolmem, unsigned long long int, 0, "outputMemory", 0, "Use predefined output buffer of this size", min(0ull), message("Using %lld bytes as output memory"))
+AddOption(registerInputMemory, bool, false, "registerInputMemory", 0, "Automatically register input memory buffers for the GPU")
 AddOption(affinity, int, -1, "cpuAffinity", 0, "Pin CPU affinity to this CPU core", min(-1), message("Setting affinity to restrict on CPU %d"))
 AddOption(fifo, bool, false, "fifoScheduler", 0, "Use FIFO realtime scheduler", message("Setting FIFO scheduler: %s"))
 AddOption(fpe, bool, true, "fpe", 0, "Trap on floating point exceptions")

GPU/GPUTracking/Standalone/standalone.cxx

@@ -329,6 +329,7 @@ int SetupReconstruction()
   devProc.globalInitMutex = configStandalone.gpuInitMutex;
   devProc.gpuDeviceOnly = configStandalone.oclGPUonly;
   devProc.memoryAllocationStrategy = configStandalone.allocationStrategy;
+  devProc.registerStandaloneInputMemory = configStandalone.registerInputMemory;
   recSet.tpcRejectionMode = configStandalone.configRec.tpcReject;
   if (configStandalone.configRec.tpcRejectThreshold != 0.f) {
     recSet.tpcRejectQPt = 1.f / configStandalone.configRec.tpcRejectThreshold;