pytorch · moksiuc · Oct 31, 2025 · Nov 13, 2025 · Nov 17, 2025 · Nov 17, 2025
diff --git a/libkineto/CMakeLists.txt b/libkineto/CMakeLists.txt
@@ -126,6 +126,10 @@ if(NOT LIBKINETO_NOCUPTI)
 endif()
 if(DEFINED LIBKINETO_NOXPUPTI AND NOT LIBKINETO_NOXPUPTI)
   list(APPEND KINETO_COMPILE_OPTIONS ${XPUPTI_BUILD_FLAG})
+  if(KINETO_BUILD_TESTS)
+    set_target_properties(kineto_base PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    set_target_properties(kineto_api PROPERTIES POSITION_INDEPENDENT_CODE ON)
+  endif()
 endif()
 if (DEFINED LIBKINETO_NOAIUPTI AND NOT LIBKINETO_NOAIUPTI)
   list(APPEND KINETO_COMPILE_OPTIONS ${AIUPTI_BUILD_FLAGS})
@@ -258,6 +262,9 @@ if(KINETO_BUILD_TESTS)
     set(BUILD_SHARED_LIBS OFF)
     add_subdirectory("${LIBKINETO_THIRDPARTY_DIR}/googletest")
     set(BUILD_SHARED_LIBS "${TMP_BUILD_SHARED_LIBS}")
+    if(DEFINED LIBKINETO_NOXPUPTI AND NOT LIBKINETO_NOXPUPTI)
+      set_property(TARGET gtest PROPERTY POSITION_INDEPENDENT_CODE ON)
+    endif()
   endif()
   enable_testing()
   add_subdirectory(test)

diff --git a/libkineto/include/ActivityType.h b/libkineto/include/ActivityType.h
@@ -41,6 +41,7 @@ enum class ActivityType {
   CUDA_PROFILER_RANGE, // CUPTI Profiler range for performance metrics
   HPU_OP, // HPU host side runtime event
   XPU_RUNTIME, // host side xpu runtime events
+  XPU_SCOPE_PROFILER, // XPUPTI Profiler scope for performance metrics
   COLLECTIVE_COMM, // collective communication
 
   // PRIVATEUSE1 Activity types are used for custom backends.

diff --git a/libkineto/include/GenericTraceActivity.h b/libkineto/include/GenericTraceActivity.h
@@ -129,6 +129,15 @@ class GenericTraceActivity : public ITraceActivity {
     return json.str();
   }
 
+  const std::vector<std::pair<std::string, std::string>> getMetadata()
+      const override {
+    std::vector<std::pair<std::string, std::string>> result;
+    for (const auto& [key, val] : metadataMap_) {
+      result.emplace_back(key, val.first);
+    }
+    return result;
+  }
+
   virtual ~GenericTraceActivity() override {}
 
   int64_t startTime{0};

diff --git a/libkineto/include/ITraceActivity.h b/libkineto/include/ITraceActivity.h
@@ -53,6 +53,10 @@ struct ITraceActivity {
   virtual const std::string getMetadataValue(const std::string& key) const {
     return "";
   }
+  virtual const std::vector<std::pair<std::string, std::string>> getMetadata()
+      const {
+    return {};
+  }
 
   static int64_t nsToUs(int64_t ns) {
     // It's important that this conversion is the same everywhere.

diff --git a/libkineto/libkineto_defs.bzl b/libkineto/libkineto_defs.bzl
@@ -38,6 +38,7 @@ def get_libkineto_xpupti_srcs(with_api = True):
         "src/plugin/xpupti/XpuptiActivityApi.cpp",
         "src/plugin/xpupti/XpuptiActivityProfiler.cpp",
         "src/plugin/xpupti/XpuptiActivityHandlers.cpp",
+        "src/plugin/xpupti/XpuptiScopeProfilerConfig.cpp",
     ] + (get_libkineto_cpu_only_srcs(with_api))
 
 def get_libkineto_aiupti_srcs(with_api = True):

diff --git a/libkineto/src/ActivityType.cpp b/libkineto/src/ActivityType.cpp
@@ -39,6 +39,7 @@ static constexpr std::array<ActivityTypeName, activityTypeCount + 1> map{
      {"cuda_profiler_range", ActivityType::CUDA_PROFILER_RANGE},
      {"hpu_op", ActivityType::HPU_OP},
      {"xpu_runtime", ActivityType::XPU_RUNTIME},
+     {"xpu_scope_profiler", ActivityType::XPU_SCOPE_PROFILER},
      {"collective_comm", ActivityType::COLLECTIVE_COMM},
      {"privateuse1_runtime", ActivityType::PRIVATEUSE1_RUNTIME},
      {"privateuse1_driver", ActivityType::PRIVATEUSE1_DRIVER},
@@ -71,8 +72,7 @@ const std::array<ActivityType, activityTypeCount> activityTypes() {
   return res;
 }
 
-const std::array<ActivityType, defaultActivityTypeCount>
-defaultActivityTypes() {
+const std::array<ActivityType, defaultActivityTypeCount> defaultActivityTypes() {
   std::array<ActivityType, defaultActivityTypeCount> res;
   for (int i = 0; i < defaultActivityTypeCount; i++) {
     res[i] = map[i].type;

diff --git a/libkineto/src/init.cpp b/libkineto/src/init.cpp
@@ -26,6 +26,7 @@
 #ifdef HAS_XPUPTI
 #include "plugin/xpupti/XpuptiActivityApi.h"
 #include "plugin/xpupti/XpuptiActivityProfiler.h"
+#include "plugin/xpupti/XpuptiScopeProfilerConfig.h"
 #endif
 #ifdef HAS_AIUPTI
 #include "plugin/aiupti/AiuptiActivityApi.h"
@@ -176,6 +177,8 @@ void libkineto_init(bool cpuOnly, bool logOnError) {
       std::make_unique<ActivityProfilerProxy>(cpuOnly, config_loader));
 
 #ifdef HAS_XPUPTI
+  XpuptiScopeProfilerConfig::registerFactory();
+
   // register xpu pti profiler
   libkineto::api().registerProfilerFactory(
       []() -> std::unique_ptr<IActivityProfiler> {
@@ -184,7 +187,7 @@ void libkineto_init(bool cpuOnly, bool logOnError) {
           std::string errPrefixMsg(
               "Fail to enable Kineto Profiler on XPU due to error code: ");
           errPrefixMsg = errPrefixMsg + std::to_string(returnCode);
-#if PTI_VERSION_MAJOR > 0 || PTI_VERSION_MINOR > 9
+#if PTI_VERSION_AT_LEAST(0, 10)
           std::string errMsg(ptiResultTypeToString(returnCode));
           throw std::runtime_error(
               errPrefixMsg + std::string(". The detailed error message is: ") +

diff --git a/libkineto/src/output_json.cpp b/libkineto/src/output_json.cpp
@@ -577,6 +577,45 @@ void ChromeTraceLogger::handleActivity(const libkineto::ITraceActivity& op) {
           toString(op.type()), op_name, device, sanitizeTid(resource),
           ts/1000, ts %1000, duration/1000, duration %1000, args);
   // clang-format on
+
+  if (op.type() == ActivityType::XPU_SCOPE_PROFILER) {
+    constexpr const std::string_view prefix = "metrics: ";
+    if (std::string_view(op_name).substr(0, prefix.size()) == prefix) {
+      for (const auto& [key, val] : op.getMetadata()) {
+        constexpr std::array<std::string_view, 2> keysExcluded = {
+            "queue", "kernel_id"};
+        if (std::none_of(
+                keysExcluded.begin(),
+                keysExcluded.end(),
+                [&key](std::string_view keyExcluded) {
+                  return keyExcluded == key;
+                })) {
+          auto timePoint = ts + duration / 2;
+          fmt::print(
+              traceOf_,
+              // clang-format off
+  R"JSON(
+  {{
+    "name": "{}",
+    "cat": "metric",
+    "ph": "C",
+    "ts": {}.{:03},
+    "pid": {},
+    "tid": {},
+    "args": {{ "value": {} }}
+  }},)JSON",
+              // clang-format on
+              key,
+              timePoint / 1000,
+              timePoint % 1000,
+              device,
+              sanitizeTid(resource),
+              val);
+        }
+      }
+    }
+  }
+
   if (op.flowId() > 0) {
     handleGenericLink(op);
   }

diff --git a/libkineto/src/plugin/xpupti/CMakeLists.txt b/libkineto/src/plugin/xpupti/CMakeLists.txt
@@ -31,8 +31,8 @@ if(TARGET Pti::pti_view)
     foreach(ver_major RANGE 0 1)
         foreach(ver_minor RANGE 0 19)
         list(APPEND pti_view_names pti_view-${ver_major}-${ver_minor})
-        endforeach()        
-    endforeach()    
+        endforeach()
+    endforeach()
 
     get_target_property(PTI_INCLUDE_DIR Pti::pti_view INTERFACE_INCLUDE_DIRECTORIES)
     find_library(PTI_VIEW_LIBRARY NAMES ${pti_view_names} PATHS "${PTI_INCLUDE_DIR}/../lib")

diff --git a/libkineto/src/plugin/xpupti/FindSYCLToolkit.cmake b/libkineto/src/plugin/xpupti/FindSYCLToolkit.cmake
@@ -105,6 +105,5 @@ if((NOT SYCL_INCLUDE_DIR) OR (NOT SYCL_LIBRARY_DIR) OR (NOT SYCL_LIBRARY))
 endif()
 
 message(DEBUG "The SYCL compiler is ${SYCL_COMPILER}")
-message(DEBUG "The SYCL Flags are ${SYCL_FLAGS}")
 
 set(SYCLTOOLKIT_FOUND True)