Write benchmark results to csv

Author: OliverRietmann

.github/workflows/standalone-benchmark.yml

@@ -33,7 +33,7 @@ jobs:
       ALIBUILD_ARCH_PREFIX: el9-x86_64/Packages
       MODULEPATH: /cvmfs/alice.cern.ch/etc/toolchain/modulefiles/el9-x86_64:/cvmfs/alice.cern.ch/el9-x86_64/Modules/modulefiles
       STANDALONE_DIR: /root/standalone
-      ARTIFACT_FILE: /root/artifact.txt
+      ARTIFACT_FILE: /root/benchmark.csv
       LD_LIBRARY_PATH: /usr/local/cuda-13.0/compat
 
     name: ${{ matrix.name }}
@@ -84,11 +84,11 @@ jobs:
           source /etc/profile.d/modules.sh
           module load ninja/fortran-v1.11.1.g9-15 Vc/1.4.5-10 boost/v1.83.0-alice2-57 fmt/11.1.2-14 CMake/v3.31.6-10 ms_gsl/4.2.1-3 Clang/v20.1.7-9 TBB/v2022.3.0-3 ROOT/v6-36-04-alice9-15 ONNXRuntime/v1.22.0-71 GLFW/3.3.2-25
           cd ${STANDALONE_DIR}
-          ${STANDALONE_DIR}/ca -e 50kHz -g --memSize 15000000000 --sync --runs 1 --debug 1 > ${ARTIFACT_FILE}
+          ${STANDALONE_DIR}/ca -e 50kHz -g --memSize 15000000000 --sync --runs 1 --debug 1 --PROCtimingCSV ${ARTIFACT_FILE}
           rm -rf ${STANDALONE_DIR}/events/50kHz ${STANDALONE_DIR}/build
 
       - name: Upload Artifact
         uses: actions/upload-artifact@v4
         with:
           name: ${{ matrix.name }}-artifact
-          path: /root/artifact.txt
+          path: /root/benchmark.csv

GPU/GPUTracking/Base/GPUReconstructionCPU.cxx

@@ -35,6 +35,7 @@
 
 #include <atomic>
 #include <ctime>
+#include <fstream>
 
 #ifndef _WIN32
 #include <unistd.h>
@@ -263,6 +264,31 @@ int32_t GPUReconstructionCPU::RunChains()
   }
   double kernelTotal = 0;
   std::vector<double> kernelStepTimes(gpudatatypes::N_RECO_STEPS, 0.);
+  std::ofstream timingCSVFile;
+  if (!GetProcessingSettings().timingCSV.empty()) {
+    bool needHeader = true;
+    {
+      std::ifstream timingCSVIn(GetProcessingSettings().timingCSV);
+      needHeader = !timingCSVIn.good() || timingCSVIn.peek() == std::ifstream::traits_type::eof();
+    }
+    timingCSVFile.open(GetProcessingSettings().timingCSV, std::ios::out | std::ios::app);
+    if (!timingCSVFile.is_open()) {
+      GPUError("Could not open timing CSV file '%s' for writing", GetProcessingSettings().timingCSV.c_str());
+    } else if (needHeader) {
+      timingCSVFile << "name,time,count,events\n";
+    }
+  }
+  auto writeCSVString = [](std::ostream& out, const std::string& s) {
+    out << '"';
+    for (char c : s) {
+      if (c == '"') {
+        out << "\"\"";
+      } else {
+        out << c;
+      }
+    }
+    out << '"';
+  };
 
   if (GetProcessingSettings().debugLevel >= 1) {
     for (uint32_t i = 0; i < mTimers.size(); i++) {
@@ -289,6 +315,10 @@ int32_t GPUReconstructionCPU::RunChains()
         snprintf(bandwidth, 256, " (%8.3f GB/s - %'14zu bytes - %'14zu per call)", mTimers[i]->memSize / time * 1e-9, mTimers[i]->memSize / mStatNEvents, mTimers[i]->memSize / mStatNEvents / mTimers[i]->count);
       }
       printf("Execution Time: Task (%c %8ux): %50s Time: %'10.0f us%s\n", type == 0 ? 'K' : 'C', mTimers[i]->count, mTimers[i]->name.c_str(), time * 1000000 / mStatNEvents, bandwidth);
+      if (timingCSVFile.is_open()) {
+        writeCSVString(timingCSVFile, mTimers[i]->name);
+        timingCSVFile << "," << (time * 1000000 / mStatNEvents) << "," << mTimers[i]->count << "," << mStatNEvents << "\n";
+      }
       if (GetProcessingSettings().resetTimers) {
         mTimers[i]->count = 0;
         mTimers[i]->memSize = 0;

GPU/GPUTracking/Definitions/GPUSettingsList.h

@@ -307,6 +307,7 @@ AddOption(debugLevel, int32_t, -1, "debug", 'd', "Set debug level (-2 = silent,
 AddOption(allocDebugLevel, int32_t, 0, "allocDebug", 0, "Some debug output for memory allocations (without messing with normal debug level)")
 AddOption(debugMask, uint32_t, (1 << 18) - 1, "debugMask", 0, "Mask for debug output dumps to file")
 AddOption(debugLogSuffix, std::string, "", "debugSuffix", 0, "Suffix for debug log files with --debug 6")
+AddOption(timingCSV, std::string, "", "", 0, "Append per-task timing rows to this CSV file")
 AddOption(serializeGPU, int8_t, 0, "", 0, "Synchronize after each kernel call (bit 1) and DMA transfer (bit 2) and identify failures")
 AddOption(recoTaskTiming, bool, 0, "", 0, "Perform summary timing after whole reconstruction tasks")
 AddOption(deterministicGPUReconstruction, int32_t, -1, "", 0, "Make CPU and GPU debug output comparable (sort / skip concurrent parts), -1 = automatic if debugLevel >= 6 or deterministic compile flag set", def(1))