Update

Author: OliverRietmann

.github/scripts/merge_runs.py

@@ -22,7 +22,7 @@
 
 data = [["name", "time", "stdev"]]
 for name, time_list in time_dict.items():
-  mean = statistics.mean(time_list[args.discard:])
+  mean = int(statistics.mean(time_list[args.discard:]))
   runs = len(time_list[args.discard:])
   stdev = 0.0 if runs == 1 else statistics.stdev(time_list[args.discard:])
   data.append([name, mean, stdev])

.github/scripts/profiler_amd.py

@@ -25,7 +25,7 @@
 data = [["name", "time", "stdev"]]
 for name, time_list in time_dict.items():
   count = len(time_list) / args.runs
-  mean = statistics.mean(time_list) * count
+  mean = int(statistics.mean(time_list) * count)
   stdev = 0 if args.runs == 1 else statistics.stdev(time_list) * count
   data.append([name, mean, stdev])
 

.github/scripts/profiler_nvidia.py

@@ -18,8 +18,8 @@
     if row:
       full_name = row[8]
       instances = int(row[2])
-      time = int(row[1]) / 1000.0
-      sigma = float(row[7]) / 1000.0
+      time = int(row[1])
+      sigma = float(row[7])
       if len(full_name) > 5 and full_name[:5] == "krnl_":
         name = full_name[5:]
         ntsi_list.append([name, time, sigma, instances])
@@ -29,7 +29,7 @@
 data = [["name", "time", "stdev"]]
 for name, time, sigma, instances in ntsi_list:
   count = instances / args.runs
-  mean = time * count
+  mean = int(time * count)
   stdev = sigma * count
   data.append([name, mean, stdev])
 

.github/workflows/standalone-benchmark.yml

@@ -21,7 +21,7 @@ jobs:
             runner: cern-nextgen-h100
             cmake_args: -DENABLE_CUDA=1 -DENABLE_HIP=0 -DCUDA_COMPUTETARGET=90
             profiler: nsys profile -o nvidia-h100
-            profiler_post: nsys stats --report cuda_gpu_kern_sum --force-export=true --format csv nvidia-h100.nsys-rep >
+            profiler_post: nsys stats --report cuda_gpu_kern_sum --timeunit us --force-export=true --format csv nvidia-h100.nsys-rep >
           - name: nvidia-l40s
             vendor: nvidia
             runner: cern-nextgen-l40s
@@ -107,7 +107,7 @@ 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}
-          ${{ matrix.profiler }} ${STANDALONE_DIR}/ca -e 50kHz -g --memSize 15000000000 --sync --debug 1 --runs 42 --runsInit 0 --PROCdebugMarkdown 1 --PROCresetTimers 1 --PROCdebugCSV /root/${BENCHMARK_CSV}
+          ${{ matrix.profiler }} ${STANDALONE_DIR}/ca -e 50kHz -g --memSize 15000000000 --sync --debug 1 --runs 42 --runsInit 2 --PROCdebugMarkdown 1 --PROCresetTimers 1 --PROCdebugCSV /root/${BENCHMARK_CSV}
           ${{ matrix.profiler_post }} /root/${PROFILER_CSV}
           rm -rf ${STANDALONE_DIR}/events/50kHz ${STANDALONE_DIR}/build
 
@@ -116,7 +116,7 @@ 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
           python3 ${GITHUB_WORKSPACE}/.github/scripts/profiler_${{ matrix.vendor }}.py --runs 42 --input /root/${PROFILER_CSV} --output /root/reduced_${PROFILER_CSV}
-          python3 ${GITHUB_WORKSPACE}/.github/scripts/merge_runs.py --discard 0 --input /root/${BENCHMARK_CSV} --output /root/reduced_${BENCHMARK_CSV}
+          python3 ${GITHUB_WORKSPACE}/.github/scripts/merge_runs.py --discard 2 --input /root/${BENCHMARK_CSV} --output /root/reduced_${BENCHMARK_CSV}
           #python3 ${GITHUB_WORKSPACE}/.github/scripts/csv_to_md.py --baseline ${STANDALONE_DIR}/baseline/${PROFILER_CSV} --current /root/${PROFILER_CSV} >> ${GITHUB_STEP_SUMMARY}
           #echo -e "\n\n" >> ${GITHUB_STEP_SUMMARY}
           #python3 ${GITHUB_WORKSPACE}/.github/scripts/csv_to_md.py --baseline ${STANDALONE_DIR}/baseline/${BENCHMARK_CSV} --current /root/${BENCHMARK_CSV} >> ${GITHUB_STEP_SUMMARY}