[SYCL] Eliminate the redundancy in pi initialization. (#1868)

Passed the implementation of pi initialization into a "call_once" function, so
that plugin initialization only needs to be executed once. After the first time,
the already initialized plugin will be returned.

Also, updated pi::initialize to return const reference to the global vector of
plugins, rather than a temporary copy of it.

Author: kholland-intel

sycl/include/CL/sycl/detail/pi.hpp

@@ -137,7 +137,7 @@ template <class To, class From> To cast(From value);
 extern std::shared_ptr<plugin> GlobalPlugin;
 
 // Performs PI one-time initialization.
-vector_class<plugin> initialize();
+const vector_class<plugin> &initialize();
 
 // Utility Functions to get Function Name for a PI Api.
 template <PiApiKind PiApiOffset> struct PiFuncInfo {};

sycl/source/detail/pi.cpp

@@ -52,6 +52,8 @@ constexpr const char *GVerStr = "sycl 1.0";
 
 namespace pi {
 
+static void initializePlugins(vector_class<plugin> *Plugins);
+
 bool XPTIInitDone = false;
 
 // Implementation of the SYCL PI API call tracing methods that use XPTI
@@ -254,8 +256,23 @@ bool trace(TraceLevel Level) {
 }
 
 // Initializes all available Plugins.
-vector_class<plugin> initialize() {
-  vector_class<plugin> Plugins;
+const vector_class<plugin> &initialize() {
+  static std::once_flag PluginsInitDone;
+  static vector_class<plugin> *Plugins = nullptr;
+
+  std::call_once(PluginsInitDone, []() {
+    // The memory for "Plugins" is intentionally leaked because the application
+    // may call into the SYCL runtime from a global destructor, and such a call
+    // could eventually call down to initialize().  Therefore, there is no safe
+    // time when "Plugins" could be deleted.
+    Plugins = new vector_class<plugin>;
+    initializePlugins(Plugins);
+  });
+
+  return *Plugins;
+}
+
+static void initializePlugins(vector_class<plugin> *Plugins) {
   vector_class<std::pair<std::string, backend>> PluginNames;
   findPlugins(PluginNames);
 
@@ -303,7 +320,7 @@ vector_class<plugin> initialize() {
       // Use the CUDA plugin as the GlobalPlugin
       GlobalPlugin = std::make_shared<plugin>(PluginInformation, backend::cuda);
     }
-    Plugins.emplace_back(plugin(PluginInformation, PluginNames[I].second));
+    Plugins->emplace_back(plugin(PluginInformation, PluginNames[I].second));
     if (trace(TraceLevel::PI_TRACE_BASIC))
       std::cerr << "SYCL_PI_TRACE[basic]: "
                 << "Plugin found and successfully loaded: "
@@ -312,7 +329,7 @@ vector_class<plugin> initialize() {
 
 #ifdef XPTI_ENABLE_INSTRUMENTATION
   if (!(xptiTraceEnabled() && !XPTIInitDone))
-    return Plugins;
+    return;
   // Not sure this is the best place to initialize the framework; SYCL runtime
   // team needs to advise on the right place, until then we piggy-back on the
   // initialization of the PI layer.
@@ -353,8 +370,6 @@ vector_class<plugin> initialize() {
       xptiMakeEvent("PI Layer", &PIPayload, xpti::trace_algorithm_event,
                     xpti_at::active, &PiInstanceNo);
 #endif
-
-  return Plugins;
 }
 
 // Report error and no return (keeps compiler from printing warnings).

sycl/source/detail/platform_impl.cpp

@@ -49,7 +49,7 @@ static bool IsBannedPlatform(platform Platform) {
 
 vector_class<platform> platform_impl::get_platforms() {
   vector_class<platform> Platforms;
-  vector_class<plugin> Plugins = RT::initialize();
+  const vector_class<plugin> &Plugins = RT::initialize();
 
   info::device_type ForcedType = detail::get_forced_type();
   for (unsigned int i = 0; i < Plugins.size(); i++) {