Fixed synchronization bug and refactoring

AliceO2Group · Feb 18, 2024 · f920549 · f920549
1 parent 45e295d
commit f920549
Show file tree

Hide file tree

Showing 8 changed files with 19 additions and 160 deletions.
diff --git a/CMakeLists.txt.orig b/CMakeLists.txt.orig
diff --git a/GPU/GPUTracking/DataCompression/GPUTPCDecompression.cxx b/GPU/GPUTracking/DataCompression/GPUTPCDecompression.cxx
@@ -94,6 +94,5 @@ void GPUTPCDecompression::RegisterMemoryAllocation()
 
 void GPUTPCDecompression::SetMaxData(const GPUTrackingInOutPointers& io)
 {
-  // mMaxNativeClustersPerBuffer = 81760;
-  mMaxNativeClustersPerBuffer = 12000;
-}
+  mMaxNativeClustersPerBuffer = mRec->GetProcessingSettings().tpcMaxAttachedClustersPerSectorRow;
+}
diff --git a/GPU/GPUTracking/DataCompression/GPUTPCDecompressionKernels.cxx b/GPU/GPUTracking/DataCompression/GPUTPCDecompressionKernels.cxx
@@ -102,8 +102,7 @@ GPUdii() void GPUTPCDecompressionKernels::decompressTrack(CompressedClusters& cm
       time = cmprClusters.timeA[trackIndex];
       pad = cmprClusters.padA[trackIndex];
     }
-    bool stored;
-    const auto cluster = decompressTrackStore(cmprClusters, clusterOffset, slice, row, pad, time, decompressor, stored);
+    const auto cluster = decompressTrackStore(cmprClusters, clusterOffset, slice, row, pad, time, decompressor);
     float y = param.tpcGeometry.LinearPad2Y(slice, row, cluster.getPad());
     float z = param.tpcGeometry.LinearTime2Z(slice, cluster.getTime());
     if (clusterIndex == 0) {
@@ -118,14 +117,16 @@ GPUdii() void GPUTPCDecompressionKernels::decompressTrack(CompressedClusters& cm
   clusterOffset += cmprClusters.nTrackClusters[trackIndex] - clusterIndex;
 }
 
-GPUdii() ClusterNative GPUTPCDecompressionKernels::decompressTrackStore(const o2::tpc::CompressedClusters& cmprClusters, const unsigned int clusterOffset, unsigned int slice, unsigned int row, unsigned int pad, unsigned int time, GPUTPCDecompression& decompressor, bool& stored)
+GPUdii() ClusterNative GPUTPCDecompressionKernels::decompressTrackStore(const o2::tpc::CompressedClusters& cmprClusters, const unsigned int clusterOffset, unsigned int slice, unsigned int row, unsigned int pad, unsigned int time, GPUTPCDecompression& decompressor)
 {
   unsigned int tmpBufferIndex = computeLinearTmpBufferIndex(slice, row, decompressor.mMaxNativeClustersPerBuffer);
   unsigned int currentClusterIndex = CAMath::AtomicAdd(decompressor.mNativeClustersIndex + (slice * GPUCA_ROW_COUNT + row), 1u);
   const ClusterNative c(time, cmprClusters.flagsA[clusterOffset], pad, cmprClusters.sigmaTimeA[clusterOffset], cmprClusters.sigmaPadA[clusterOffset], cmprClusters.qMaxA[clusterOffset], cmprClusters.qTotA[clusterOffset]);
-  stored = currentClusterIndex < decompressor.mMaxNativeClustersPerBuffer;
-  if (stored) {
+  if (currentClusterIndex < decompressor.mMaxNativeClustersPerBuffer) {
     decompressor.mTmpNativeClusters[tmpBufferIndex + currentClusterIndex] = c;
+  } else {
+    decompressor.raiseError(GPUErrors::ERROR_DECOMPRESSION_ATTACHED_CLUSTER_OVERFLOW, slice * 1000 + row, currentClusterIndex, decompressor.mMaxNativeClustersPerBuffer);
+    CAMath::AtomicExch(decompressor.mNativeClustersIndex + (slice * GPUCA_ROW_COUNT + row), decompressor.mMaxNativeClustersPerBuffer);
   }
   return c;
 }

diff --git a/GPU/GPUTracking/DataCompression/GPUTPCDecompressionKernels.h b/GPU/GPUTracking/DataCompression/GPUTPCDecompressionKernels.h
@@ -47,7 +47,7 @@ class GPUTPCDecompressionKernels : public GPUKernelTemplate
   template <int iKernel = defaultKernel>
   GPUd() static void Thread(int nBlocks, int nThreads, int iBlock, int iThread, GPUsharedref() GPUSharedMemory& smem, processorType& GPUrestrict() processors);
   GPUd() static void decompressTrack(o2::tpc::CompressedClusters& cmprClusters, const GPUParam& param, const unsigned int maxTime, const unsigned int trackIndex, unsigned int& clusterOffset, GPUTPCDecompression& decompressor);
-  GPUdi() static o2::tpc::ClusterNative decompressTrackStore(const o2::tpc::CompressedClusters& cmprClusters, const unsigned int clusterOffset, unsigned int slice, unsigned int row, unsigned int pad, unsigned int time, GPUTPCDecompression& decompressor, bool& stored);
+  GPUdi() static o2::tpc::ClusterNative decompressTrackStore(const o2::tpc::CompressedClusters& cmprClusters, const unsigned int clusterOffset, unsigned int slice, unsigned int row, unsigned int pad, unsigned int time, GPUTPCDecompression& decompressor);
   GPUdi() static void decompressHits(const o2::tpc::CompressedClusters& cmprClusters, const unsigned int start, const unsigned int end, o2::tpc::ClusterNative* clusterNativeBuffer);
 
   GPUd() static unsigned int computeLinearTmpBufferIndex(unsigned int slice, unsigned int row, unsigned int maxClustersPerBuffer)

diff --git a/GPU/GPUTracking/Definitions/GPUSettingsList.h b/GPU/GPUTracking/Definitions/GPUSettingsList.h
@@ -260,6 +260,7 @@ AddOption(throttleAlarms, bool, false, "", 0, "Throttle rate at which alarms are
 AddOption(outputSanityCheck, bool, false, "", 0, "Run some simple sanity checks finding errors in the output")
 AddOption(tpcSingleSector, int, -1, "", 0, "Restrict TPC processing to a single sector")
 AddOption(tpcDownscaledEdx, unsigned char, 0, "", 0, "If != 0, downscale dEdx processing (if enabled) to x %")
+AddOption(tpcMaxAttachedClustersPerSectorRow, unsigned int, 20000, "", 0, "Maximum number of TPC attached clusters which can be decoded per SectorRow")
 AddVariable(eventDisplay, GPUCA_NAMESPACE::gpu::GPUDisplayFrontendInterface*, nullptr)
 AddSubConfig(GPUSettingsProcessingRTC, rtc)
 AddSubConfig(GPUSettingsProcessingParam, param)

diff --git a/GPU/GPUTracking/Global/GPUChainTrackingCompression.cxx b/GPU/GPUTracking/Global/GPUChainTrackingCompression.cxx
@@ -298,7 +298,9 @@ int GPUChainTracking::RunTPCDecompression()
   processors()->ioPtrs.clustersNative = mInputsHost->mPclusterNativeAccess;
 
   runKernel<GPUTPCDecompressionKernels, GPUTPCDecompressionKernels::step1unattached>(GetGridAutoStep(inputStream, RecoStep::TPCDecompression), krnlRunRangeNone, krnlEventNone);
-  GPUMemCpy(RecoStep::TPCDecompression, mInputsHost->mPclusterNativeOutput, mInputsShadow->mPclusterNativeBuffer, sizeof(mInputsShadow->mPclusterNativeBuffer[0]) * mIOPtrs.clustersNative->nClustersTotal, inputStream, false);
+  SynchronizeStream(inputStream);
+  GPUMemCpy(RecoStep::TPCDecompression, (void*)mInputsHost->mPclusterNativeOutput, (void*)mInputsShadow->mPclusterNativeBuffer, sizeof(mInputsShadow->mPclusterNativeBuffer[0]) * mIOPtrs.clustersNative->nClustersTotal, inputStream, false);
+
   mIOPtrs.clustersNative = mClusterNativeAccess.get();
 #endif
   return 0;

diff --git a/GPU/GPUTracking/Global/GPUErrorCodes.h b/GPU/GPUTracking/Global/GPUErrorCodes.h
@@ -40,11 +40,12 @@ GPUCA_ERROR_CODE(20, ERROR_CF_PEAK_OVERFLOW, Sector, Value, Max)
 GPUCA_ERROR_CODE(21, ERROR_CF_CLUSTER_OVERFLOW, Sector, Value, Max)
 GPUCA_ERROR_CODE(22, ERROR_CF_ROW_CLUSTER_OVERFLOW, SectorRow, Value, Max)
 GPUCA_ERROR_CODE(23, ERROR_CF_GLOBAL_CLUSTER_OVERFLOW, SectorRow, Value, Max)
-GPUCA_ERROR_CODE(24, MAX_OVERFLOW_ERROR_NUMBER) // Overflow errors are detected as errno <= MAX_OVERFLOW_ERROR_NUMBER
+GPUCA_ERROR_CODE(24, ERROR_DECOMPRESSION_ATTACHED_CLUSTER_OVERFLOW, SectorRow, Value, Max)
+GPUCA_ERROR_CODE(25, MAX_OVERFLOW_ERROR_NUMBER) // Overflow errors are detected as errno <= MAX_OVERFLOW_ERROR_NUMBER
 
-GPUCA_ERROR_CODE(25, ERROR_TPCZS_INVALID_ROW, SectorRow)                                  // Data from invalid row is skipped
-GPUCA_ERROR_CODE(26, ERROR_TPCZS_INVALID_NADC, SectorCRU, SamplesInPage, SamplesWritten)  // Invalid number of ADC samples in header, existing samples were decoded
-GPUCA_ERROR_CODE(27, ERROR_TPCZS_INCOMPLETE_HBF, SectorCRU, PacketCount, NextPacketCount) // Part of HBF is missing, decoding incomplete
-GPUCA_ERROR_CODE(28, ERROR_TPCZS_INVALID_OFFSET, SectorEndpoint, Value, Expected)         // Raw page is skipped since it contains invalid payload offset
+GPUCA_ERROR_CODE(26, ERROR_TPCZS_INVALID_ROW, SectorRow)                                  // Data from invalid row is skipped
+GPUCA_ERROR_CODE(27, ERROR_TPCZS_INVALID_NADC, SectorCRU, SamplesInPage, SamplesWritten)  // Invalid number of ADC samples in header, existing samples were decoded
+GPUCA_ERROR_CODE(28, ERROR_TPCZS_INCOMPLETE_HBF, SectorCRU, PacketCount, NextPacketCount) // Part of HBF is missing, decoding incomplete
+GPUCA_ERROR_CODE(29, ERROR_TPCZS_INVALID_OFFSET, SectorEndpoint, Value, Expected)         // Raw page is skipped since it contains invalid payload offset
 
 // #define GPUCA_CHECK_TPCZS_CORRUPTION
diff --git a/GPU/GPUTracking/Standalone/Benchmark/standalone.cxx b/GPU/GPUTracking/Standalone/Benchmark/standalone.cxx
@@ -950,7 +950,6 @@ int main(int argc, char** argv)
       printf("Error unregistering memory\n");
     }
   }
-  // exit(0);
   rec->Exit();
 
   if (!configStandalone.noprompt) {