Skip to content

Commit 3f977d1

Browse files
ggalgocziggalgoczi
committed
tests: GPU-vs-G4 hit-comparison CI test for the dRICH GDML geometry
Adds tests/test_drich_gdml.sh (run on the [gpu] CI runner alongside the other test_*.sh) that runs the standalone dRICH example through Geant4 CPU and Opticks GPU side-by-side and asserts the sensor-hit counts agree within NSIGMA*sqrt(G+C) (Poisson). Builds the example via its CMakeLists if no binary is installed. SKIPs (benign) without a GPU or if the rare G4 navigator hang times out every seed. Wired into build-pull-request.yaml's GPU test step.
1 parent 00e676d commit 3f977d1

2 files changed

Lines changed: 129 additions & 0 deletions

File tree

‎.github/workflows/build-pull-request.yaml‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -187,6 +187,7 @@ jobs:
187187
docker run --rm --gpus 'device=1' ${{ env.IMAGE_NAME }}:${{ env.IMAGE_TAG }} tests/test_GPUPhotonFileSource.sh
188188
docker run --rm --gpus 'device=1' ${{ env.IMAGE_NAME }}:${{ env.IMAGE_TAG }} tests/test_GPUPhotonSource_8x8SiPM.sh
189189
docker run --rm --gpus 'device=1' ${{ env.IMAGE_NAME }}:${{ env.IMAGE_TAG }} tests/test_wavelength_shifting.sh
190+
docker run --rm --gpus 'device=1' ${{ env.IMAGE_NAME }}:${{ env.IMAGE_TAG }} tests/test_drich_gdml.sh
190191
191192
- name: Cleanup local test image
192193
if: ${{ success() }}

‎tests/test_drich_gdml.sh‎

Lines changed: 128 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,128 @@
1+
#!/bin/bash
2+
#
3+
# test_drich_gdml.sh
4+
# ===================
5+
# End-to-end test: Opticks GPU vs Geant4 CPU for the standalone (no-DD4hep)
6+
# dRICH GDML geometry (examples/drich). The driver runs both transports in one
7+
# process and prints "gpu_hits=N cpu_hits=M" per event; this wraps it with a
8+
# statistical pass/fail on the GPU/CPU sensor-hit agreement.
9+
#
10+
# Geometry: examples/drich/drich_ag02.gdml (single-sector dRICH, 0.2 mm airgap)
11+
#
12+
# Comparison uses Poisson statistics (CLAUDE.md house rule): for counts
13+
# G(pu) and C(pu) the difference error is sqrt(G+C), so the test passes when
14+
# |G - C| <= NSIGMA * sqrt(G + C)
15+
# i.e. GPU and CPU agree within NSIGMA standard deviations. The validated
16+
# dRICH agreement is ~1 sigma; NSIGMA defaults to 5 for a robust CI gate that
17+
# still flags the gross divergences the geometry fixes addressed (0.6x..1.25x).
18+
#
19+
# The G4 CPU optical navigator has a rare (~1-in-150k photon) hang on the
20+
# aerogel/airgap coincident face (a Geant4 11.04 bug, not Opticks). We run with
21+
# a step watchdog + per-try timeout and retry a few seeds; if every try times
22+
# out we SKIP (exit 0) rather than fail, keeping CI non-flaky.
23+
#
24+
# Env knobs (all optional): DRICH_BIN MULT NEVENTS NSIGMA TIMEOUT_S SEEDS
25+
# CUDA_VISIBLE_DEVICES OPTICKS_PREFIX
26+
#
27+
# Usage: ./tests/test_drich_gdml.sh
28+
# Exit: 0 = PASS or SKIP, 1 = FAIL
29+
#
30+
31+
set -u
32+
33+
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
34+
REPO_DIR="$(cd "$SCRIPT_DIR/.." && pwd)"
35+
GDML="$REPO_DIR/examples/drich/drich_ag02.gdml"
36+
37+
MULT="${MULT:-20}"
38+
NEVENTS="${NEVENTS:-2}"
39+
NSIGMA="${NSIGMA:-5}"
40+
TIMEOUT_S="${TIMEOUT_S:-300}"
41+
SEEDS="${SEEDS:-12345 23456 34567}"
42+
43+
echo "=============================================="
44+
echo " dRICH GDML test: Opticks GPU vs Geant4 CPU"
45+
echo "=============================================="
46+
47+
# --- GPU availability: this is a GPU test, SKIP (benign) where there is none.
48+
if ! command -v nvidia-smi >/dev/null 2>&1 || ! nvidia-smi -L >/dev/null 2>&1; then
49+
echo "SKIPPED: no GPU available"
50+
exit 0
51+
fi
52+
53+
if [ ! -f "$GDML" ]; then
54+
echo "SKIPPED: $GDML not found"
55+
exit 0
56+
fi
57+
58+
source /opt/eic-opticks/eic-opticks-env.sh 2>/dev/null || true
59+
60+
# --- Obtain the driver: prefer an installed/explicit binary, else build the
61+
# example against the installed library (also exercises examples/drich/CMakeLists.txt).
62+
BIN="${DRICH_BIN:-}"
63+
if [ -z "$BIN" ] && command -v drich_gdml_main >/dev/null 2>&1; then
64+
BIN="drich_gdml_main"
65+
fi
66+
if [ -z "$BIN" ]; then
67+
BUILD_DIR="${TMPDIR:-/tmp}/drich_ci_build"
68+
LOG="${TMPDIR:-/tmp}/drich_ci_build.log"
69+
echo "[build] configuring + building examples/drich ..."
70+
if cmake -S "$REPO_DIR/examples/drich" -B "$BUILD_DIR" \
71+
-DCMAKE_PREFIX_PATH="${OPTICKS_PREFIX:-/opt/eic-opticks}" >"$LOG" 2>&1 \
72+
&& cmake --build "$BUILD_DIR" >>"$LOG" 2>&1; then
73+
BIN="$BUILD_DIR/drich_gdml_main"
74+
echo "[build] OK -> $BIN"
75+
else
76+
echo "FAIL: could not build examples/drich"
77+
tail -25 "$LOG"
78+
exit 1
79+
fi
80+
fi
81+
82+
echo " driver: $BIN"
83+
echo " geometry: $GDML"
84+
echo " config: MULT=$MULT NEVENTS=$NEVENTS NSIGMA=$NSIGMA TIMEOUT_S=$TIMEOUT_S"
85+
echo ""
86+
87+
# --- Run: G4 CPU + Opticks GPU side-by-side. Retry seeds to survive the rare
88+
# CPU navigator hang; first clean run (gpu>0 && cpu>0) wins.
89+
GPU=0
90+
CPU=0
91+
USED_SEED=""
92+
for SEED in $SEEDS; do
93+
echo "[run] seed=$SEED ..."
94+
OUT=$(QBND_FILTER_POINT=1 KILL_OPTICAL=0 \
95+
GDML_FILE="$GDML" OPTICKS_MAX_SLOT="${OPTICKS_MAX_SLOT:-2000000}" \
96+
MULT="$MULT" SEED="$SEED" NEVENTS="$NEVENTS" PHI_DEG=30 ETA=2.0 \
97+
MAX_STEP_WATCHDOG="${MAX_STEP_WATCHDOG:-100000}" \
98+
timeout --signal=KILL "${TIMEOUT_S}s" "$BIN" 2>&1) || true
99+
g=$(echo "$OUT" | grep -oE "gpu_hits=[0-9]+" | grep -oE "[0-9]+" | awk '{s+=$1} END{print s+0}')
100+
c=$(echo "$OUT" | grep -oE "cpu_hits=[0-9]+" | grep -oE "[0-9]+" | awk '{s+=$1} END{print s+0}')
101+
echo " gpu_hits=$g cpu_hits=$c"
102+
if [ "$g" -gt 0 ] && [ "$c" -gt 0 ]; then
103+
GPU=$g; CPU=$c; USED_SEED=$SEED
104+
break
105+
fi
106+
echo " (no clean result this seed -- likely CPU navigator hang/timeout, retrying)"
107+
done
108+
109+
if [ -z "$USED_SEED" ]; then
110+
echo "SKIPPED: no clean G4+GPU run obtained in $(echo $SEEDS | wc -w) tries"
111+
echo " (known Geant4 11.04 navigator hang at the aerogel/airgap face)"
112+
exit 0
113+
fi
114+
115+
# --- Compare with Poisson (sqrt(N)) statistics.
116+
echo ""
117+
echo "[compare] seed=$USED_SEED GPU=$GPU CPU=$CPU"
118+
awk -v g="$GPU" -v c="$CPU" -v ns="$NSIGMA" 'BEGIN{
119+
diff = g - c; adiff = (diff<0)? -diff : diff;
120+
sig = adiff / sqrt(g + c);
121+
ratio= g / c;
122+
tol = ns * sqrt(g + c);
123+
printf " ratio GPU/CPU = %.4f\n", ratio;
124+
printf " |GPU-CPU| = %d ; sqrt(G+C) = %.1f ; significance = %.2f sigma (tol %d-sigma => %.0f)\n", adiff, sqrt(g+c), sig, ns, tol;
125+
if (sig <= ns) { print "PASS: GPU and CPU hit counts agree within tolerance"; exit 0 }
126+
else { print "FAIL: GPU vs CPU hit-count difference exceeds tolerance"; exit 1 }
127+
}'
128+
exit $?

0 commit comments

Comments
 (0)