[xpupti] Fix CPU->GPU flow links via activity type, not an API-name allowlist

libkineto/src/plugin/xpupti/XpuptiActivityHandlers.cpp

@@ -214,8 +214,7 @@ void XpuptiActivityProfilerSession::handleRuntimeKernelMemcpyMemsetActivities(
   if constexpr (handleRuntimeActivities) {
     trace_activity->device = activity->_process_id;
     trace_activity->resource = activity->_thread_id;
-    trace_activity->flow.start =
-        (correlateRuntimeOps_.count(trace_activity->name()) > 0);
+    trace_activity->flow.start = startsFlow(activityType);
   } else {
     trace_activity->device = getDeviceIdxFromUUID(activity->_device_uuid);
     trace_activity->resource = activity->_sycl_queue_id;

libkineto/src/plugin/xpupti/XpuptiActivityProfilerSession.cpp

@@ -19,28 +19,17 @@
 
 namespace KINETO_NAMESPACE {
 
-using namespace std::literals::string_view_literals;
-
 uint32_t XpuptiActivityProfilerSession::iterationCount_ = 0;
 std::vector<DeviceUUIDsT> XpuptiActivityProfilerSession::deviceUUIDs_ = {};
-std::unordered_set<std::string_view>
-    XpuptiActivityProfilerSession::correlateRuntimeOps_ = {
-        "piextUSMEnqueueFill"sv,
-        "urEnqueueUSMFill"sv,
-        "piextUSMEnqueueFill2D"sv,
-        "urEnqueueUSMFill2D"sv,
-        "piextUSMEnqueueMemcpy"sv,
-        "urEnqueueUSMMemcpy"sv,
-        "piextUSMEnqueueMemset"sv,
-        "piextUSMEnqueueMemcpy2D"sv,
-        "urEnqueueUSMMemcpy2D"sv,
-        "piextUSMEnqueueMemset2D"sv,
-        "piEnqueueKernelLaunch"sv,
-        "urEnqueueKernelLaunch"sv,
-        "piextEnqueueKernelLaunchCustom"sv,
-        "urEnqueueKernelLaunchCustomExp"sv,
-        "piextEnqueueCooperativeKernelLaunch"sv,
-        "urEnqueueCooperativeKernelLaunchExp"sv};
+
+bool XpuptiActivityProfilerSession::startsFlow(ActivityType activityType) {
+  // Only host runtime records start the CPU->GPU flow. The runtime view is
+  // already filtered to work-submitting APIs via
+  // ptiViewEnableRuntimeApiClass(PTI_API_CLASS_GPU_OPERATION_CORE). Driver
+  // (XPU_DRIVER) records share the same correlation id as the runtime record
+  // and must not also start a flow, or the trace gets a duplicate flow start.
+  return activityType == ActivityType::XPU_RUNTIME;
+}
 
 // =========== Session Constructor ============= //
 XpuptiActivityProfilerSession::XpuptiActivityProfilerSession(

libkineto/src/plugin/xpupti/XpuptiActivityProfilerSession.h

@@ -18,7 +18,6 @@
 #include <memory>
 #include <set>
 #include <unordered_map>
-#include <unordered_set>
 #include <vector>
 
 namespace KINETO_NAMESPACE {
@@ -61,6 +60,12 @@ class XpuptiActivityProfilerSession : public libkineto::IActivityProfilerSession
   void pushUserCorrelationId(uint64_t id) override;
   void popUserCorrelationId() override;
 
+  // Whether a runtime/driver record starts a CPU->GPU flow arrow. Only host
+  // runtime (XPU_RUNTIME) records do; driver (XPU_DRIVER) records share the
+  // same correlation id and would otherwise create a duplicate flow start.
+  // Static so it can be unit-tested without real hardware.
+  static bool startsFlow(ActivityType activityType);
+
  private:
   void checkTimestampOrder(const ITraceActivity* act1);
   void removeCorrelatedPtiActivities(const ITraceActivity* act1);
@@ -94,7 +99,6 @@ class XpuptiActivityProfilerSession : public libkineto::IActivityProfilerSession
  protected:
   static uint32_t iterationCount_;
   static std::vector<DeviceUUIDsT> deviceUUIDs_;
-  static std::unordered_set<std::string_view> correlateRuntimeOps_;
 
   int64_t captureWindowStartTime_{0};
   int64_t captureWindowEndTime_{0};

libkineto/test/xpupti/CMakeLists.txt

@@ -19,6 +19,11 @@ add_executable(XpuptiScopeProfilerConfigTest XpuptiScopeProfilerConfigTest.cpp)
 target_link_libraries(XpuptiScopeProfilerConfigTest PRIVATE ${LINK_LIBRARIES})
 gtest_discover_tests(XpuptiScopeProfilerConfigTest)
 
+# Hardware-free test for the CPU->GPU flow-link allowlist.
+add_executable(XpuptiFlowCorrelationTest XpuptiFlowCorrelationTest.cpp)
+target_link_libraries(XpuptiFlowCorrelationTest PRIVATE ${LINK_LIBRARIES})
+gtest_discover_tests(XpuptiFlowCorrelationTest)
+
 include(ExternalProject)
 
 function(make_test test_file)

libkineto/test/xpupti/XpuptiFlowCorrelationTest.cpp

@@ -0,0 +1,41 @@
+//==============================================================
+// Copyright (C) Intel Corporation
+//
+// SPDX-License-Identifier: MIT
+// =============================================================
+
+#include "src/plugin/xpupti/XpuptiActivityProfilerSession.h"
+
+#include <gtest/gtest.h>
+
+namespace KN = KINETO_NAMESPACE;
+
+// Exercises the predicate that decides which records start a CPU->GPU ("ac2g")
+// flow arrow. Runs without a GPU: only the static predicate is queried, no PTI
+// collection is performed.
+
+TEST(XpuptiFlowCorrelationTest, RuntimeRecordsStartFlow) {
+  // Host runtime records (kernel launches, memcpy/fill enqueues) are the source
+  // of the CPU->GPU arrow. The runtime view is filtered to work-submitting APIs
+  // by ptiViewEnableRuntimeApiClass(PTI_API_CLASS_GPU_OPERATION_CORE).
+  EXPECT_TRUE(KN::XpuptiActivityProfilerSession::startsFlow(
+      KN::ActivityType::XPU_RUNTIME));
+}
+
+TEST(XpuptiFlowCorrelationTest, DriverRecordsDoNotStartFlow) {
+  // Driver (ze*) records share the runtime record's correlation id; if they
+  // also started a flow the trace would have a duplicate flow start for that
+  // id (Perfetto flow_duplicate_id). They must only be flow ends.
+  EXPECT_FALSE(KN::XpuptiActivityProfilerSession::startsFlow(
+      KN::ActivityType::XPU_DRIVER));
+}
+
+TEST(XpuptiFlowCorrelationTest, DeviceRecordsDoNotStartFlow) {
+  // GPU-side activities are flow destinations, never sources.
+  EXPECT_FALSE(KN::XpuptiActivityProfilerSession::startsFlow(
+      KN::ActivityType::CONCURRENT_KERNEL));
+  EXPECT_FALSE(KN::XpuptiActivityProfilerSession::startsFlow(
+      KN::ActivityType::GPU_MEMCPY));
+  EXPECT_FALSE(KN::XpuptiActivityProfilerSession::startsFlow(
+      KN::ActivityType::GPU_MEMSET));
+}