Skip to content

Commit 4088d78

Browse files
OliverRietmannOliverRietmann
committed
Test CI
1 parent f5212ca commit 4088d78

1 file changed

Lines changed: 9 additions & 101 deletions

File tree

‎.github/workflows/standalone-benchmark.yml‎

Lines changed: 9 additions & 101 deletions
Original file line numberDiff line numberDiff line change
@@ -9,54 +9,19 @@ on:
99

1010
jobs:
1111
benchmark:
12-
runs-on: ${{ matrix.runner }}
12+
runs-on: cern-nextgen-mi300x
1313
container: registry.cern.ch/alisw/slc9-gpu-builder@sha256:ea3443f9dfbc770e4b4bce0d1a9ecc0b7a7c16e9f76e416b796d170877220820
14-
strategy:
15-
fail-fast: false
16-
matrix:
17-
name: [cpu, nvidia-h100, nvidia-l40s, amd-mi300x, amd-w7900]
18-
include:
19-
- name: cpu
20-
runner: cern-nextgen-mi300x
21-
cmake_args: -DENABLE_CUDA=0 -DENABLE_HIP=0
22-
profiler_runs: 42
23-
standalone_runs: 42
24-
cpu_gpu: "-c"
25-
- name: nvidia-h100
26-
runner: cern-nextgen-h100
27-
cmake_args: -DENABLE_CUDA=1 -DENABLE_HIP=0 -DCUDA_COMPUTETARGET=90
28-
profiler_runs: 21
29-
standalone_runs: 42
30-
cpu_gpu: "-g --memSize 20000000000"
31-
- name: nvidia-l40s
32-
runner: cern-nextgen-l40s
33-
cmake_args: -DENABLE_CUDA=1 -DENABLE_HIP=0 -DCUDA_COMPUTETARGET=89
34-
profiler_runs: 42
35-
standalone_runs: 42
36-
cpu_gpu: "-g --memSize 20000000000"
37-
- name: amd-mi300x
38-
runner: cern-nextgen-mi300x
39-
cmake_args: -DENABLE_CUDA=0 -DENABLE_HIP=1 -DHIP_AMDGPUTARGET=gfx942
40-
profiler_runs: 42
41-
standalone_runs: 42
42-
cpu_gpu: "-g --memSize 20000000000"
43-
- name: amd-w7900
44-
runner: cern-nextgen-w7900
45-
cmake_args: -DENABLE_CUDA=0 -DENABLE_HIP=1 -DHIP_AMDGPUTARGET=gfx1100
46-
profiler_runs: 42
47-
standalone_runs: 42
48-
cpu_gpu: "-g --memSize 20000000000"
4914
env:
5015
WORK_DIR: /cvmfs/alice.cern.ch
5116
ALIBUILD_ARCH_PREFIX: el9-x86_64/Packages
5217
MODULEPATH: /cvmfs/alice.cern.ch/etc/toolchain/modulefiles/el9-x86_64:/cvmfs/alice.cern.ch/el9-x86_64/Modules/modulefiles
5318
STANDALONE_DIR: /root/standalone
54-
BENCHMARK_CSV: standalone_${{ matrix.name }}.csv
55-
PROFILER_CSV: profiler_${{ matrix.name }}.csv
56-
TIMING_CA: ./ca -e 50kHz ${{ matrix.cpu_gpu }} --seed 0 --sync --runsInit 0 --PROCresetTimers 1 --PROCdebugMarkdown 1 --debug 1 # Add --runs 42 for benchmark runs
19+
BENCHMARK_CSV: standalone_cpu.csv
20+
PROFILER_CSV: profiler_cpu.csv
21+
TIMING_CA: ./ca -e 50kHz -c --seed 0 --sync --runsInit 0 --PROCresetTimers 1 --PROCdebugMarkdown 1 --debug 1 # Add --runs 42 for benchmark runs
5722
LD_LIBRARY_PATH: /usr/local/cuda-13.0/compat
5823

59-
name: ${{ matrix.name }}
24+
name: cpu
6025
steps:
6126
- name: Checkout Repository
6227
uses: actions/checkout@v6
@@ -80,84 +45,27 @@ jobs:
8045
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
8146

8247
mkdir -p ${STANDALONE_DIR}
83-
cmake -B ${STANDALONE_DIR}/build ${{ matrix.cmake_args }} -DENABLE_OPENCL=0 -DGPUCA_BUILD_EVENT_DISPLAY=0 -DGPUCA_DETERMINISTIC_MODE=${DETERMINISTIC_MODE} -DCMAKE_INSTALL_PREFIX=${STANDALONE_DIR} ${GITHUB_WORKSPACE}/GPU/GPUTracking/Standalone/
48+
cmake -B ${STANDALONE_DIR}/build -DCMAKE_INSTALL_PREFIX=${STANDALONE_DIR} ${GITHUB_WORKSPACE}/GPU/GPUTracking/Standalone/
8449
cmake --build ${STANDALONE_DIR}/build --target install -j 8
85-
env:
86-
DETERMINISTIC_MODE: GPU
8750

8851
- name: Test Track Reconstruction
8952
run: |
9053
source /etc/profile.d/modules.sh
9154
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
9255
cd ${STANDALONE_DIR}
93-
${STANDALONE_DIR}/ca -e o2-simple ${{ matrix.cpu_gpu }} --seed 0 --sync --runs 1 --RTCenable --PROCdeterministicGPUReconstruction 1 --RTCoptConstexpr 1 --RTCoptSpecialCode 1 --debug 6
56+
${STANDALONE_DIR}/ca -e o2-simple -c --seed 0 --sync --runs 1 --RTCenable --PROCdeterministicGPUReconstruction 1 --RTCoptConstexpr 1 --RTCoptSpecialCode 1 --debug 6
9457
cmp ${STANDALONE_DIR}/*.out
95-
rm -rf ${STANDALONE_DIR}/*.out ${STANDALONE_DIR}/events/o2-simple ${STANDALONE_DIR}/build
96-
97-
- name: Build Non-Deterministic
98-
run: *build
99-
env:
100-
DETERMINISTIC_MODE: OFF
10158
10259
- name: Benchmark Track Reconstruction
10360
run: |
10461
source /etc/profile.d/modules.sh
10562
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
10663
cd ${STANDALONE_DIR}
107-
${TIMING_CA} --runs ${{ matrix.standalone_runs }} --PROCdebugCSV /root/${BENCHMARK_CSV}
64+
${TIMING_CA} --runs 42 --PROCdebugCSV /root/${BENCHMARK_CSV}
10865
python3 ${GITHUB_WORKSPACE}/.github/scripts/profiler_standalone.py --discard 0 --input /root/${BENCHMARK_CSV} --output /root/summary_${BENCHMARK_CSV}
109-
110-
- name: Profiler - Nsight Compute
111-
if: ${{ matrix.name == 'nvidia-h100' }}
112-
run: |
113-
dnf install -y cuda-nsight-compute-13-1
114-
source /etc/profile.d/modules.sh
115-
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
116-
cd ${STANDALONE_DIR}
117-
ncu --set none --metrics gpu__time_duration.avg --export ${{ matrix.name }} --clock-control none --force-overwrite ${TIMING_CA} --runs ${{ matrix.profiler_runs }} # Generates ${{ matrix.name }}.ncu-rep
118-
ncu --import ${STANDALONE_DIR}/${{ matrix.name }}.ncu-rep --print-units base --csv > /root/${PROFILER_CSV}
119-
rm -rf ${STANDALONE_DIR}/events/50kHz ${STANDALONE_DIR}/build
120-
python3 ${GITHUB_WORKSPACE}/.github/scripts/profiler_ncu.py --input /root/${PROFILER_CSV} --output /root/summary_${PROFILER_CSV}
121-
122-
- name: Profiler - Nsight Systems
123-
if: ${{ matrix.name == 'nvidia-l40s' }}
124-
run: |
125-
dnf config-manager --add-repo "https://developer.download.nvidia.com/devtools/repos/rhel$(source /etc/os-release; echo ${VERSION_ID%%.*})/$(rpm --eval '%{_arch}' | sed s/aarch/arm/)/"
126-
dnf install --nogpgcheck -y nsight-systems-cli-2026.2.1
127-
source /etc/profile.d/modules.sh
128-
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
129-
cd ${STANDALONE_DIR}
130-
nsys profile -o ${{ matrix.name }} ${TIMING_CA} --runs ${{ matrix.profiler_runs }} # Generates ${{ matrix.name }}.nsys-rep
131-
nsys stats --report cuda_gpu_kern_sum --timeunit usec --force-export=true --format csv ${{ matrix.name }}.nsys-rep > /root/${PROFILER_CSV}
132-
rm -rf ${STANDALONE_DIR}/events/50kHz ${STANDALONE_DIR}/build
133-
python3 ${GITHUB_WORKSPACE}/.github/scripts/profiler_nsys.py --input /root/${PROFILER_CSV} --output /root/summary_${PROFILER_CSV}
134-
135-
- name: Profiler - rocprofv2
136-
if: ${{ matrix.name == 'amd-mi300x' || matrix.name == 'amd-w7900' }}
137-
run: |
138-
source /etc/profile.d/modules.sh
139-
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
140-
cd ${STANDALONE_DIR}
141-
rocprofv2 --output-directory /root --output-file-name ${{ matrix.name }} ${TIMING_CA} --runs ${{ matrix.standalone_runs }} # Generates results_${{ matrix.name }}.csv
142-
rm -rf ${STANDALONE_DIR}/events/50kHz ${STANDALONE_DIR}/build
143-
mv /root/results_${{ matrix.name }}.csv /root/${PROFILER_CSV}
144-
python3 ${GITHUB_WORKSPACE}/.github/scripts/profiler_rocprofv2.py --input /root/${PROFILER_CSV} --output /root/summary_${PROFILER_CSV}
14566
14667
- name: Upload Artifact
14768
uses: actions/upload-artifact@v6
14869
with:
149-
name: ${{ matrix.name }}-artifact
70+
name: cpu-artifact
15071
path: "/root/*.csv"
151-
152-
- name: Display table on GitHub web
153-
run: |
154-
source /etc/profile.d/modules.sh
155-
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
156-
mkdir -p ${STANDALONE_DIR}/baseline
157-
curl -fL --retry 3 -o ${STANDALONE_DIR}/baseline/summary_${PROFILER_CSV} https://cernbox.cern.ch/remote.php/dav/public-files/SfYXgQOHFga2w75/baseline/summary_${PROFILER_CSV}
158-
curl -fL --retry 3 -o ${STANDALONE_DIR}/baseline/summary_${BENCHMARK_CSV} https://cernbox.cern.ch/remote.php/dav/public-files/SfYXgQOHFga2w75/baseline/summary_${BENCHMARK_CSV}
159-
python3 ${GITHUB_WORKSPACE}/.github/scripts/csv_to_md.py --runs ${{ matrix.profiler_runs }} --baseline ${STANDALONE_DIR}/baseline/summary_${PROFILER_CSV} --current /root/summary_${PROFILER_CSV} >> ${GITHUB_STEP_SUMMARY}
160-
echo -e "\n\n" >> ${GITHUB_STEP_SUMMARY}
161-
python3 ${GITHUB_WORKSPACE}/.github/scripts/csv_to_md.py --runs ${{ matrix.standalone_runs }} --baseline ${STANDALONE_DIR}/baseline/summary_${BENCHMARK_CSV} --current /root/summary_${BENCHMARK_CSV} >> ${GITHUB_STEP_SUMMARY}
162-
rm -rf ${STANDALONE_DIR}/baseline
163-
if: ${{ matrix.name != 'cpu' }}

0 commit comments

Comments
 (0)