[XPTI][INFRA] Fixes minor standalone build and code issues (#18452)

This PR is first of many XPTI changes to address the performance issues
that have been reported.

(1) xpti::Framework::instance() - testing the use of std::call_once
against current implementation
(2) Added cleanup for default stream "xpti.framework"
(3) Switched to EmHash instead of std::unordered_map that improve
notification performance by 10%

Most of the performance overheads are from xptiRegisterObject() and
xptiAddMetadata(). This requires changes in the SYCL runtime an will be
addressed in a subsequent PR

---------

Signed-off-by: Vasanth Tovinkere <[email protected]>

Author: tovinkere

xptifw/CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 3.20.0)
 
-set(XPTI_VERSION 1.0.0)
+set(XPTI_VERSION 1.0.1)
 
 project (xptifw VERSION "${XPTI_VERSION}" LANGUAGES CXX)
 set(CMAKE_CXX_STANDARD 17)
@@ -74,8 +74,8 @@ endif()
 
 if (XPTI_BUILD_SAMPLES)
   add_subdirectory(samples/basic_collector)
-  add_subdirectory(samples/syclpi_collector)
   add_subdirectory(samples/sycl_perf_collector)
+  add_subdirectory(samples/syclur_collector)
   add_subdirectory(basic_test)
 endif()
 

xptifw/CMakeLists.txt.in

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 2.8.2)
+cmake_minimum_required(VERSION 3.20.0)
 
 project(googletest-download NONE)
 

xptifw/samples/sycl_perf_collector/sycl_perf_collector.cpp

@@ -41,6 +41,7 @@ event_instances_t *GRecordsInProgress = nullptr;
 xpti::utils::timer::measurement_t GMeasure;
 
 constexpr const char *GStreamBasic = "sycl";
+constexpr const char *GStreamPerf = "sycl.perf";
 constexpr const char *GStreamPI = "ur.call";
 constexpr const char *GStreamMemory = "sycl.experimental.mem_alloc";
 constexpr const char *GStreamL0 = "sycl.experimental.level_zero.call";
@@ -291,12 +292,14 @@ XPTI_CALLBACK_API void xptiTraceInit(unsigned int major_version,
     // In order to determine if the environment variable contains valid stream
     // names, a catalog of all streams must be check against to validate the
     // strings
-    if (ShowVerboseOutput && StreamsToMonitor) {
-      std::cout << "Monitoring streams: ";
+    if (StreamsToMonitor) {
+      if (ShowVerboseOutput)
+        std::cout << "Monitoring streams: ";
 
       auto streams = D.decode(StreamsToMonitor);
       for (auto &s : streams) {
-        std::cout << s << ", ";
+        if (ShowVerboseOutput)
+          std::cout << s << ", ";
         GAllStreams.add(s.c_str());
       }
       if (ShowVerboseOutput)
@@ -394,17 +397,17 @@ XPTI_CALLBACK_API void xptiTraceInit(unsigned int major_version,
     InitStreams = false;
   } // First time inititalization complete
 
-  if (ShowVerboseOutput)
-    std::cout << "Initializing stream: " << stream_name << "\n";
-
   // Post initialization: Once the needed data structures are created, we
   // register callbacks to the streams requsted by the end-user;
   //
   // Check == TRUE if the stream has been requested by end-user
   //
   auto Check = GAllStreams.check(stream_name);
-  if (Check)
+  if (Check) {
     GStreams->add(stream_name);
+    if (ShowVerboseOutput)
+      std::cout << "Initializing stream: " << stream_name << "\n";
+  }
 
   if (std::string(GStreamBasic) == stream_name && Check) {
     auto StreamID = xptiRegisterStream(stream_name);
@@ -436,10 +439,10 @@ XPTI_CALLBACK_API void xptiTraceInit(unsigned int major_version,
     // Handles SelfNotification events
     xptiRegisterCallback(StreamID,
                          (uint16_t)xpti::trace_point_type_t::function_begin,
-                         syclPiCallback);
+                         syclURCallback);
     xptiRegisterCallback(StreamID,
                          (uint16_t)xpti::trace_point_type_t::function_end,
-                         syclPiCallback);
+                         syclURCallback);
   } else if (std::string(GStreamMemory) == stream_name && Check) {
     auto StreamID = xptiRegisterStream(stream_name);
     xptiRegisterCallback(StreamID,
@@ -468,6 +471,14 @@ XPTI_CALLBACK_API void xptiTraceInit(unsigned int major_version,
     xptiRegisterCallback(StreamID,
                          (uint16_t)xpti::trace_point_type_t::function_end,
                          syclURCallback);
+  } else if (std::string(GStreamPerf) == stream_name && Check) {
+    auto StreamID = xptiRegisterStream(stream_name);
+    xptiRegisterCallback(StreamID,
+                         (uint16_t)xpti::trace_point_type_t::function_begin,
+                         syclURCallback);
+    xptiRegisterCallback(StreamID,
+                         (uint16_t)xpti::trace_point_type_t::function_end,
+                         syclURCallback);
   } else if (std::string(GStreamL0) == stream_name && Check) {
     auto StreamID = xptiRegisterStream(stream_name);
     xptiRegisterCallback(StreamID,
@@ -515,7 +526,7 @@ std::once_flag GFinalize, GCompaction;
 // terminated sending events in its stream
 XPTI_CALLBACK_API void xptiTraceFinish(const char *stream_name) {
   {
-    if (ShowVerboseOutput)
+    if (ShowVerboseOutput && GStreams->check(stream_name))
       std::cout << "Unregistering stream: " << stream_name << "\n";
     std::lock_guard<std::mutex> _{GStreamMutex};
     // Filter the streams to what is requested by the user that intersects with

xptifw/src/CMakeLists.txt

@@ -1,4 +1,5 @@
 # Fetch third party headers
+include(FetchContent)
 set(EMHASH_REPO https://github.com/ktprime/emhash)
 message(STATUS "Will fetch emhash from ${EMHASH_REPO}")
 FetchContent_Declare(emhash-headers

xptifw/src/xpti_trace_framework.cpp

@@ -82,8 +82,16 @@ static thread_local xpti_tracepoint_t *g_tls_tracepoint_scope_data;
 /// Default stream for the trace framework, if none is provided by the user.
 constexpr const char *g_default_stream = "xpti.framework";
 
+/// @brief Flag to ensure the default stream is initialized only once.
+/// @details Used with `std::call_once` to initialize the default stream in a
+/// thread-safe manner.
 static std::once_flag g_initialize_default_stream_flag;
 
+/// @brief Flag to ensure the default stream is finalized only once.
+/// @details Used with `std::call_once` to finalize the default stream in a
+/// thread-safe manner.
+static std::once_flag g_finalize_default_stream_flag;
+
 namespace xpti {
 /// @var env_subscribers
 /// @brief A constant character pointer initialized with the string
@@ -1379,15 +1387,15 @@ class Notifications {
   /// @brief Maps a trace type to its associated callback entries.
   /// @details This unordered map uses a uint16_t as the key to represent the
   /// trace point type, and cb_entries_t to store the associated callbacks.
-  using cb_t = std::unordered_map<uint16_t, cb_entries_t>;
+  using cb_t = emhash7::HashMap<uint16_t, cb_entries_t>;
 
   /// @typedef stream_cb_t
   /// @brief Maps a stream ID to its corresponding callbacks for different
   /// trace types
   /// @details This unordered map uses a uint16_t as the key for the stream
   /// ID, and cb_t to map the stream to registered callbacks for each trace
   /// type
-  using stream_cb_t = std::unordered_map<uint16_t, cb_t>;
+  using stream_cb_t = emhash7::HashMap<uint16_t, cb_t>;
 
   /// @typedef statistics_t
   /// @brief Keeps track of statistics, typically counts, associated with
@@ -1396,13 +1404,13 @@ class Notifications {
   /// the type of statistical data and usually not defined by default. To
   /// enable it, XPTI_STATISTICS has to be defined while compiling the
   /// frmaework library.
-  using statistics_t = std::unordered_map<uint16_t, uint64_t>;
+  using statistics_t = emhash7::HashMap<uint16_t, uint64_t>;
   /// @typedef trace_flags_t
   /// @brief Maps an trace type to a boolean flag indicating its state.
   /// @details This unordered map uses a uint16_t as the key for the trace
   /// type, and a boolean value to indicate whether callbacks are registered
   /// for this trace type (e.g., registered or unregisterted/no callback).
-  using trace_flags_t = std::unordered_map<uint16_t, bool>;
+  using trace_flags_t = emhash7::HashMap<uint16_t, bool>;
 
   /// @typedef stream_flags_t
   /// @brief Maps a stream ID to its corresponding trace flags for different
@@ -1411,7 +1419,7 @@ class Notifications {
   /// and trace_flags_t to map the trace type to their boolean that indiciates
   /// whether a callback has been registered for this trace type in the given
   /// stream.
-  using stream_flags_t = std::unordered_map<uint8_t, trace_flags_t>;
+  using stream_flags_t = emhash7::HashMap<uint8_t, trace_flags_t>;
 
   /// @brief Registers a callback function for a specific trace type and stream
   /// ID.
@@ -2290,6 +2298,9 @@ class Framework {
   }
 
   static Framework &instance() {
+    // Using std::call_once has the same overhead as the original approach
+    // std::call_once(g_initialize_framework_flag,
+    //               [&]() { MInstance = new Framework(); });
     Framework *TmpFramework = MInstance.load(std::memory_order_relaxed);
     std::atomic_thread_fence(std::memory_order_acquire);
     if (TmpFramework == nullptr) {
@@ -2309,6 +2320,11 @@ class Framework {
   friend void ::xptiFrameworkFinalize();
 
   static void release() {
+    // Using std::call_once has the same overhead as the original approach
+    // std::call_once(g_release_framework_flag, [&]() {
+    //   delete MInstance;
+    //   MInstance = nullptr;
+    // });
     Framework *TmpFramework = MInstance.load(std::memory_order_relaxed);
     MInstance.store(nullptr, std::memory_order_relaxed);
     delete TmpFramework;
@@ -2482,7 +2498,10 @@ XPTI_EXPORT_API xpti::result_t xptiInitialize(const char *Stream, uint32_t maj,
 /// when the stream is no longer in use.
 
 XPTI_EXPORT_API void xptiFinalize(const char *Stream) {
-  xpti::Framework::instance().finalizeStream(Stream);
+  auto &FW = xpti::Framework::instance();
+  std::call_once(g_finalize_default_stream_flag,
+                 [&]() { FW.finalizeStream(g_default_stream); });
+  FW.finalizeStream(Stream);
 }
 
 /// @brief Retrieves the 64-bit universal ID for the current scope, if published

xptifw/unit_test/CMakeLists.txt

@@ -7,7 +7,7 @@ include_directories(${XPTI_DIR}/include)
 if (NOT DEFINED LLVM_EXTERNAL_XPTIFW_SOURCE_DIR)
   message(STATUS "Building XPTI outside of LLVM Project...")
   # Download and unpack googletest at configure time
-  configure_file(../CMakeLists.txt.in googletest-download/CMakeLists.txt)
+  configure_file(${XPTIFW_DIR}/CMakeLists.txt.in googletest-download/CMakeLists.txt)
   execute_process(COMMAND ${CMAKE_COMMAND} -G "${CMAKE_GENERATOR}" .
     RESULT_VARIABLE result
     WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/googletest-download )