From 727727445e096cc83283edd589595c4359f088fa Mon Sep 17 00:00:00 2001
From: yoshifuminakamura <nakamura@riken.jp>
Date: Wed, 16 Sep 2026 08:50:03 +0900
Subject: [PATCH] Enable SBD GPU profile flow by default

Signed-off-by: yoshifuminakamura <nakamura@riken.jp>
---
 programs/sbd/README.md                |  7 ++++---
 programs/sbd/profile.sh               | 17 +++++++++++++++--
 scripts/tests/test_sbd_ncu_profile.sh | 13 +++++++++----
 3 files changed, 28 insertions(+), 9 deletions(-)

diff --git a/programs/sbd/README.md b/programs/sbd/README.md
index 136443b..8e92c79 100644
--- a/programs/sbd/README.md
+++ b/programs/sbd/README.md
@@ -5,9 +5,10 @@ diagonalization benchmark from `github.com/r-ccs-cms/sbd`.
 
 ## Optional NSYS / NCU profiling
 
-SBD keeps ordinary benchmark timing unprofiled. Set `BK_PROFILER=ncu` or
-`SBD_PROFILER_TOOL=ncu` for a separate profiling run; `run.sh` maps that
-request to the SBD NCU flow. The profile flow runs rank-0 host
+SBD keeps ordinary benchmark timing unprofiled, then collects a separate
+profiling run on supported NVIDIA GPU systems. Set `BK_SBD_NCU_PROFILE=false`
+or `SBD_PROFILER_TOOL=none` to skip the additional profile flow. The profile
+flow runs rank-0 host
 `nsys profile --trace=cuda --sample=none`, exports the CUDA kernel summary,
 generates `results/sbd_kernel_discovery.json` and `results/sbd_ncu_plan.json`,
 and then uses that plan for rank-0 Nsight Compute acquisition when
diff --git a/programs/sbd/profile.sh b/programs/sbd/profile.sh
index 0eddf5b..4e8b877 100644
--- a/programs/sbd/profile.sh
+++ b/programs/sbd/profile.sh
@@ -13,16 +13,29 @@ sbd_ncu_profile_mode() {
   printf '%s\n' "${BK_SBD_NCU_PROFILE_MODE:-discovery}"
 }
 
+sbd_supports_ncu_profile() {
+  case "$1" in
+    RIKYU|RC_DGXSP|RC_GH200) return 0 ;;
+    *) return 1 ;;
+  esac
+}
+
 sbd_configure_ncu_profile_from_run_env() {
   local system_name="$1"
   local profiler_tool
   local profiler_level
 
-  profiler_tool=$(bk_resolve_profiler_tool none SBD_PROFILER_TOOL) || return 1
-  if [ -z "$profiler_tool" ]; then
+  if ! sbd_supports_ncu_profile "$system_name"; then
     return 0
   fi
 
+  profiler_tool=$(bk_resolve_profiler_tool ncu SBD_PROFILER_TOOL) || return 1
+  if [ -z "$profiler_tool" ]; then
+    if [ -z "${BK_SBD_NCU_PROFILE:-}" ]; then
+      export BK_SBD_NCU_PROFILE=false
+    fi
+    return 0
+  fi
   if [ "$profiler_tool" != "ncu" ]; then
     echo "SBD ${system_name}: only ncu is supported for separate profile acquisition." >&2
     return 1
diff --git a/scripts/tests/test_sbd_ncu_profile.sh b/scripts/tests/test_sbd_ncu_profile.sh
index 6396117..e614b89 100644
--- a/scripts/tests/test_sbd_ncu_profile.sh
+++ b/scripts/tests/test_sbd_ncu_profile.sh
@@ -142,15 +142,20 @@ chmod +x "${FAKE_BIN}/mpirun" "${FAKE_BIN}/nsys" "${FAKE_BIN}/ncu" "${RUN_DIR}/d
 export PATH="${FAKE_BIN}:${PATH}"
 
 (
-  export BK_PROFILER=ncu
+  unset BK_PROFILER BK_PROFILER_LEVEL BK_SBD_NCU_PROFILE BK_SBD_NCU_PROFILER_LEVEL SBD_PROFILER_TOOL
+  sbd_configure_ncu_profile_from_run_env RC_FX700
+  test -z "${BK_SBD_NCU_PROFILE:-}"
+)
+
+(
+  unset BK_PROFILER BK_PROFILER_LEVEL BK_SBD_NCU_PROFILE BK_SBD_NCU_PROFILER_LEVEL
   export SBD_PROFILER_TOOL=none
-  unset BK_SBD_NCU_PROFILE BK_SBD_NCU_PROFILER_LEVEL
   sbd_configure_ncu_profile_from_run_env RIKYU
-  test -z "${BK_SBD_NCU_PROFILE:-}"
+  test "${BK_SBD_NCU_PROFILE:-}" = "false"
 )
 
 pushd "${RUN_DIR}" >/dev/null
-export BK_PROFILER=ncu
+unset BK_PROFILER SBD_PROFILER_TOOL BK_SBD_NCU_PROFILE BK_SBD_NCU_PROFILER_LEVEL
 export BK_SBD_NCU_PROFILE_MODE=discovery
 export BK_SBD_NCU_PLAN_TOP_K=5
 export BK_PROFILER_LEVEL=single