GPU: Do tfTime extrapolation also for wall time

Author: davidrohr

GPU/GPUTracking/Base/GPUReconstruction.h

@@ -238,6 +238,7 @@ class GPUReconstruction
   // Support / Debugging
   virtual void PrintKernelOccupancies() {}
   double GetStatKernelTime() { return mStatKernelTime; }
+  double GetStatWallTime() { return mStatWallTime; }
 
  protected:
   GPUReconstruction(const GPUSettingsProcessing& cfg); // Constructor
@@ -326,6 +327,7 @@ class GPUReconstruction
   unsigned int mStatNEvents = 0;
   unsigned int mNEventsProcessed = 0;
   double mStatKernelTime = 0.;
+  double mStatWallTime = 0.;
 
   int mMaxThreads = 0; // Maximum number of threads that may be running, on CPU or GPU
   int mThreadId = -1;  // Thread ID that is valid for the local CUDA context

GPU/GPUTracking/Base/GPUReconstructionCPU.cxx

@@ -198,6 +198,7 @@ int GPUReconstructionCPU::RunChains()
   }
   timerTotal.Stop();
 
+  mStatWallTime = (timerTotal.GetElapsedTime() * 1000000. / mStatNEvents);
   if (GetDeviceProcessingSettings().debugLevel >= 1) {
     double kernelTotal = 0;
     std::vector<double> kernelStepTimes(N_RECO_STEPS);
@@ -254,10 +255,10 @@ int GPUReconstructionCPU::RunChains()
       }
     }
     mStatKernelTime = kernelTotal * 1000000 / mStatNEvents;
-    printf("Execution Time: Total   : %50s Time: %'10d us\n", "Total kernel time", (int)mStatKernelTime);
-    printf("Execution Time: Total   : %50s Time: %'10d us\n", "Total time", (int)(timerTotal.GetElapsedTime() * 1000000. / mStatNEvents));
+    printf("Execution Time: Total   : %50s Time: %'10d us\n", "Total Kernel", (int)mStatKernelTime);
+    printf("Execution Time: Total   : %50s Time: %'10d us\n", "Total Wall", (int)mStatWallTime);
   } else if (GetDeviceProcessingSettings().debugLevel >= 0) {
-    printf("Total Time: %'d us\n", (int)(timerTotal.GetElapsedTime() * 1000000 / mStatNEvents));
+    printf("Total Wall Time: %'d us\n", (int)mStatWallTime);
   }
   if (mDeviceProcessingSettings.resetTimers) {
     mStatNEvents = 0;

GPU/GPUTracking/Standalone/standalone.cxx

@@ -201,10 +201,6 @@ int ReadConfiguration(int argc, char** argv)
     printf("Can only produce QA pdf output when input files are specified!\n");
     return 1;
   }
-  if (configStandalone.timeFrameTime && configStandalone.DebugLevel == 0) {
-    printf("tfTime needs debug >= 1 currently\n");
-    return 1;
-  }
   if (configStandalone.eventDisplay) {
     configStandalone.noprompt = 1;
   }
@@ -792,15 +788,15 @@ int main(int argc, char** argv)
           double nClusters = chainTracking->GetTPCMerger().NMaxClusters();
           if (nClusters > 0) {
             double nClsPerTF = 550000. * 1138.3;
-            double timePerTF = rec->GetStatKernelTime() / 1000000. * nClsPerTF / nClusters;
+            double timePerTF = (configStandalone.DebugLevel ? rec->GetStatKernelTime() : rec->GetStatWallTime()) / 1000000. * nClsPerTF / nClusters;
             double nGPUsReq = timePerTF / 0.02277;
             char stat[1024];
             snprintf(stat, 1024, "Sync phase: %.2f sec per 256 orbit TF, %.1f GPUs required", timePerTF, nGPUsReq);
             if (configStandalone.testSyncAsync) {
-              timePerTF = recAsync->GetStatKernelTime() / 1000000. * nClsPerTF / nClusters;
+              timePerTF = (configStandalone.DebugLevel ? recAsync->GetStatKernelTime() : recAsync->GetStatWallTime()) / 1000000. * nClsPerTF / nClusters;
               snprintf(stat + strlen(stat), 1024 - strlen(stat), " - Async phase: %f sec per TF", timePerTF);
             }
-            printf("%s (Extrapolated from %d clusters to %d)\n", stat, (int)nClusters, (int)nClsPerTF);
+            printf("%s (Measured %s time - Extrapolated from %d clusters to %d)\n", stat, configStandalone.DebugLevel ? "kernel" : "wall", (int)nClusters, (int)nClsPerTF);
           }
         }
       }