Add profiler_rocm.py

Author: OliverRietmann

.github/scripts/profiler_rocm.py

@@ -0,0 +1,34 @@
+import argparse
+import csv
+import math
+import statistics
+
+parser = argparse.ArgumentParser()
+parser.add_argument('-r', '--runs', type=int, required=True, help='Number of runs')
+parser.add_argument('-i', '--input', required=True, help='Input CSV file')
+parser.add_argument('-o', '--output', required=True, help='Output CSV file')
+args = parser.parse_args()
+
+time_dict = dict({})
+with open(args.input) as csv_file:
+  csv_reader = csv.reader(csv_file)
+  next(csv_reader)
+  for row in csv_reader:
+    full_name = row[13]
+    time = (int(row[15]) - int(row[14])) / 1000.0
+    if len(full_name) > 5 and full_name[:5] == "krnl_":
+      name = full_name[5:-3]
+      if name in time_dict.keys():
+        time_dict[name].append(time)
+      else:
+        time_dict[name] = [time]
+
+data = [["name", "time", "stdev"]]
+for name, time_list in time_dict.items():
+  mean = math.fsum(time_list) / args.runs
+  stdev = 0 if args.runs == 1 else statistics.stdev(time_list) * len(time_list) / args.runs
+  data.append([name, mean, stdev])
+
+with open(args.output, 'w') as csv_file:
+  csv_writer = csv.writer(csv_file)
+  csv_writer.writerows(data)

.github/workflows/standalone-benchmark.yml

@@ -27,17 +27,18 @@ jobs:
           - name: amd-mi300x
             runner: cern-nextgen-mi300x
             cmake_args: -DENABLE_CUDA=0 -DENABLE_HIP=1 -DHIP_AMDGPUTARGET=gfx942
-            profiler: rocprofv2 --basenames --kernel-trace --output-directory /root --output-file-name amd-mi300x --plugin perfetto
+            profiler: rocprofv2 --basenames --output-directory /root --output-file-name amd-mi300x
           - name: amd-w7900
             runner: cern-nextgen-w7900
             cmake_args: -DENABLE_CUDA=0 -DENABLE_HIP=1 -DHIP_AMDGPUTARGET=gfx1100
-            profiler: rocprofv2 --basenames --kernel-trace --output-directory /root --output-file-name amd-w7900 --plugin perfetto
+            profiler: rocprofv2 --basenames --output-directory /root --output-file-name amd-w7900
     env:
       WORK_DIR: /cvmfs/alice.cern.ch
       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
       BENCHMARK_CSV: /root/${{ matrix.name }}.csv
+      PROFILER_CSV: /root/results_${{ matrix.name }}.csv
       LD_LIBRARY_PATH: /usr/local/cuda-13.0/compat
 
     name: ${{ matrix.name }}
@@ -91,14 +92,14 @@ 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 --debug 1 --runs 12 --runsInit 2 --PROCresetTimers 1 --PROCtimingCSV ${BENCHMARK_CSV}
-          ${{ matrix.profiler }} ${STANDALONE_DIR}/ca -e 50kHz -g --memSize 15000000000 --sync --runs 10 --runsInit 0 --PROCresetTimers 1
+          ${{ matrix.profiler }} ${STANDALONE_DIR}/ca -e 50kHz -g --memSize 15000000000 --sync --debug 1 --runs 12 --runsInit 2 --PROCresetTimers 1 --PROCtimingCSV ${BENCHMARK_CSV}
           rm -rf ${STANDALONE_DIR}/events/50kHz ${STANDALONE_DIR}/build
 
       - name: Display table on GitHub web
         run: |
           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_rocm.py --runs 12 --input ${PROFILER_CSV} --output ${PROFILER_CSV}
           python3 ${GITHUB_WORKSPACE}/.github/scripts/merge_runs.py --discard 2 --input ${BENCHMARK_CSV} --output ${BENCHMARK_CSV}
           python3 ${GITHUB_WORKSPACE}/.github/scripts/csv_to_md.py --baseline ${STANDALONE_DIR}/baseline/${{ matrix.name }}.csv --current ${BENCHMARK_CSV} >> ${GITHUB_STEP_SUMMARY}
           rm -rf ${STANDALONE_DIR}/baseline
@@ -107,6 +108,4 @@ jobs:
         uses: actions/upload-artifact@v6
         with:
           name: ${{ matrix.name }}-artifact
-          path: |
-            /root/*.csv
-            /root/*.pftrace
+          path: "/root/*.csv"