GPU: fix count of CompressionGatherKernel in timing output

Author: davidrohr

GPU/GPUTracking/Base/GPUReconstructionCPU.cxx

@@ -357,15 +357,14 @@ GPUReconstructionCPU::timerMeta* GPUReconstructionCPU::insertTimer(uint32_t id,
   return retVal;
 }
 
-GPUReconstructionCPU::timerMeta* GPUReconstructionCPU::getTimerById(uint32_t id)
+GPUReconstructionCPU::timerMeta* GPUReconstructionCPU::getTimerById(uint32_t id, bool increment)
 {
   timerMeta* retVal = nullptr;
   while (timerFlag.test_and_set()) {
-    ;
   }
   if (mTimers.size() > id && mTimers[id]) {
     retVal = mTimers[id].get();
-    retVal->count++;
+    retVal->count += increment;
   }
   timerFlag.clear();
   return retVal;

GPU/GPUTracking/Base/GPUReconstructionCPU.h

@@ -177,7 +177,7 @@ class GPUReconstructionCPU : public GPUReconstructionKernels<GPUReconstructionCP
   RecoStepTimerMeta mTimersRecoSteps[GPUDataTypes::N_RECO_STEPS];
   HighResTimer timerTotal;
   template <class T, int32_t I = 0>
-  HighResTimer& getKernelTimer(RecoStep step, int32_t num = 0, size_t addMemorySize = 0);
+  HighResTimer& getKernelTimer(RecoStep step, int32_t num = 0, size_t addMemorySize = 0, bool increment = true);
   template <class T, int32_t J = -1>
   HighResTimer& getTimer(const char* name, int32_t num = -1);
 
@@ -186,7 +186,7 @@ class GPUReconstructionCPU : public GPUReconstructionKernels<GPUReconstructionCP
  private:
   size_t TransferMemoryResourcesHelper(GPUProcessor* proc, int32_t stream, bool all, bool toGPU);
   uint32_t getNextTimerId();
-  timerMeta* getTimerById(uint32_t id);
+  timerMeta* getTimerById(uint32_t id, bool increment = true);
   timerMeta* insertTimer(uint32_t id, std::string&& name, int32_t J, int32_t num, int32_t type, RecoStep step);
 };
 
@@ -272,10 +272,10 @@ inline void GPUReconstructionCPU::AddGPUEvents(T*& events)
 }
 
 template <class T, int32_t I>
-HighResTimer& GPUReconstructionCPU::getKernelTimer(RecoStep step, int32_t num, size_t addMemorySize)
+HighResTimer& GPUReconstructionCPU::getKernelTimer(RecoStep step, int32_t num, size_t addMemorySize, bool increment)
 {
   static int32_t id = getNextTimerId();
-  timerMeta* timer = getTimerById(id);
+  timerMeta* timer = getTimerById(id, increment);
   if (timer == nullptr) {
     timer = insertTimer(id, GetKernelName<T, I>(), -1, NSLICES, 0, step);
   }

GPU/GPUTracking/Global/GPUChain.h

@@ -190,9 +190,9 @@ class GPUChain
   }
 
   template <class T, int32_t I = 0>
-  HighResTimer& getKernelTimer(RecoStep step, int32_t num = 0, size_t addMemorySize = 0)
+  HighResTimer& getKernelTimer(RecoStep step, int32_t num = 0, size_t addMemorySize = 0, bool increment = true)
   {
-    return mRec->getKernelTimer<T, I>(step, num, addMemorySize);
+    return mRec->getKernelTimer<T, I>(step, num, addMemorySize, increment);
   }
   template <class T, int32_t J = -1>
   HighResTimer& getTimer(const char* name, int32_t num = -1)

GPU/GPUTracking/Global/GPUChainTrackingCompression.cxx

@@ -100,24 +100,24 @@ int32_t GPUChainTracking::RunTPCCompression()
     switch (ProcessingSettings().tpcCompressionGatherModeKernel) {
       case 0:
         runKernel<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::unbuffered>(GetGridBlkStep(nBlocksDefault, outputStream, RecoStep::TPCCompression));
-        getKernelTimer<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::unbuffered>(RecoStep::TPCCompression, 0, outputSize);
+        getKernelTimer<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::unbuffered>(RecoStep::TPCCompression, 0, outputSize, false);
         break;
       case 1:
         runKernel<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::buffered32>(GetGridBlkStep(nBlocksDefault, outputStream, RecoStep::TPCCompression));
-        getKernelTimer<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::buffered32>(RecoStep::TPCCompression, 0, outputSize);
+        getKernelTimer<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::buffered32>(RecoStep::TPCCompression, 0, outputSize, false);
         break;
       case 2:
         runKernel<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::buffered64>(GetGridBlkStep(nBlocksDefault, outputStream, RecoStep::TPCCompression));
-        getKernelTimer<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::buffered64>(RecoStep::TPCCompression, 0, outputSize);
+        getKernelTimer<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::buffered64>(RecoStep::TPCCompression, 0, outputSize, false);
         break;
       case 3:
         runKernel<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::buffered128>(GetGridBlkStep(nBlocksDefault, outputStream, RecoStep::TPCCompression));
-        getKernelTimer<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::buffered128>(RecoStep::TPCCompression, 0, outputSize);
+        getKernelTimer<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::buffered128>(RecoStep::TPCCompression, 0, outputSize, false);
         break;
       case 4:
         static_assert((nBlocksMulti & 1) && nBlocksMulti >= 3);
         runKernel<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::multiBlock>(GetGridBlkStep(nBlocksMulti, outputStream, RecoStep::TPCCompression));
-        getKernelTimer<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::multiBlock>(RecoStep::TPCCompression, 0, outputSize);
+        getKernelTimer<GPUTPCCompressionGatherKernels, GPUTPCCompressionGatherKernels::multiBlock>(RecoStep::TPCCompression, 0, outputSize, false);
         break;
       default:
         GPUError("Invalid compression kernel %d selected.", (int32_t)ProcessingSettings().tpcCompressionGatherModeKernel);