From 4fc62d97730327bc6e192a4183d0717df33aa017 Mon Sep 17 00:00:00 2001
From: Ernst Hellbar <ernst.hellbar@cern.ch>
Date: Wed, 4 Jun 2025 09:53:33 +0200
Subject: [PATCH] dpl-workflow.sh: use different MI100 serialization workaround
 with better performance

---
 prodtests/full-system-test/dpl-workflow.sh | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/prodtests/full-system-test/dpl-workflow.sh b/prodtests/full-system-test/dpl-workflow.sh
index dd575099857f4..729def1264f5e 100755
--- a/prodtests/full-system-test/dpl-workflow.sh
+++ b/prodtests/full-system-test/dpl-workflow.sh
@@ -270,7 +270,8 @@ if [[ $GPUTYPE == "HIP" ]]; then
     TIMESLICEOFFSET=$(($GPU_FIRST_ID + ($NUMAGPUIDS != 0 ? ($NGPUS * $NUMAID) : 0)))
     GPU_CONFIG+=" --environment \"ROCR_VISIBLE_DEVICES={timeslice${TIMESLICEOFFSET}}\""
   fi
-  [[ $EPNSYNCMODE == 1 || ! -z ${OPTIMIZED_PARALLEL_ASYNC:-} ]] && [[ ${EPN_NODE_MI100:-} == "1" ]] && [[ ${DISABLE_MI100_SERIALIZATION:-0} != 1 ]] && GPU_CONFIG_KEY+="GPU_proc.serializeGPU=3;"
+  # serialization workaround for MI100 nodes: remove it again if the problem will be fixed in ROCm, then also remove the DISABLE_MI100_SERIALIZATION flag in the O2DPG parse script
+  [[ $EPNSYNCMODE == 1 || ! -z ${OPTIMIZED_PARALLEL_ASYNC:-} ]] && [[ ${EPN_NODE_MI100:-} == "1" ]] && [[ ${DISABLE_MI100_SERIALIZATION:-0} != 1 ]] && GPU_CONFIG_KEY+="GPU_proc.amdMI100SerializationWorkaround=1;"
   #export HSA_TOOLS_LIB=/opt/rocm/lib/librocm-debug-agent.so.2
 else
   GPU_CONFIG_KEY+="GPU_proc.deviceNum=-2;"