[SYCL] Make ur::getAdapter return raw adapter pointer instead of shared_ptr (#19102)

Currently, `ur::getAdapter` returns a shared pointer to the adapter.
However, that's unnecessary.
This PR makes `ur::getAdapter` return a raw ptr instead and let global
handler manage lifetime of the adapter object.

Author: uditagarwal97

sycl/source/detail/adapter.hpp

@@ -239,8 +239,6 @@ class Adapter {
   UrFuncPtrMapT UrFuncPtrs;
 }; // class Adapter
 
-using AdapterPtr = std::shared_ptr<Adapter>;
-
 } // namespace detail
 } // namespace _V1
 } // namespace sycl

sycl/source/detail/context_impl.hpp

@@ -366,8 +366,7 @@ void GetCapabilitiesIntersectionSet(const std::vector<sycl::device> &Devices,
 // convenient to be able to reference them without extra `detail::`.
 inline auto get_ur_handles(sycl::detail::context_impl &Ctx) {
   ur_context_handle_t urCtx = Ctx.getHandleRef();
-  const sycl::detail::Adapter *Adapter = Ctx.getAdapter().get();
-  return std::tuple{urCtx, Adapter};
+  return std::tuple{urCtx, Ctx.getAdapter()};
 }
 inline auto get_ur_handles(const sycl::context &syclContext) {
   return get_ur_handles(*sycl::detail::getSyclObjImpl(syclContext));

sycl/source/detail/global_handler.cpp

@@ -230,8 +230,8 @@ std::mutex &GlobalHandler::getFilterMutex() {
   return FilterMutex;
 }
 
-std::vector<AdapterPtr> &GlobalHandler::getAdapters() {
-  static std::vector<AdapterPtr> &adapters = getOrCreate(MAdapters);
+std::vector<Adapter *> &GlobalHandler::getAdapters() {
+  static std::vector<Adapter *> &adapters = getOrCreate(MAdapters);
   enableOnCrashStackPrinting();
   return adapters;
 }
@@ -314,6 +314,7 @@ void GlobalHandler::unloadAdapters() {
   if (MAdapters.Inst) {
     for (const auto &Adapter : getAdapters()) {
       Adapter->release();
+      delete Adapter;
     }
   }
 
@@ -387,6 +388,10 @@ void shutdown_late() {
   Handler->MScheduler.Inst.reset(nullptr);
   Handler->MProgramManager.Inst.reset(nullptr);
 
+  // Cache stores handles to the adapter, so clear it before
+  // releasing adapters.
+  Handler->MKernelNameBasedCaches.Inst.reset(nullptr);
+
   // Clear the adapters and reset the instance if it was there.
   Handler->unloadAdapters();
   if (Handler->MAdapters.Inst)

sycl/source/detail/global_handler.hpp

@@ -30,7 +30,6 @@ class ThreadPool;
 struct KernelNameBasedCacheT;
 
 using ContextImplPtr = std::shared_ptr<context_impl>;
-using AdapterPtr = std::shared_ptr<Adapter>;
 
 /// Wrapper class for global data structures with non-trivial destructors.
 ///
@@ -71,7 +70,7 @@ class GlobalHandler {
   std::mutex &getPlatformToDefaultContextCacheMutex();
   std::mutex &getPlatformMapMutex();
   std::mutex &getFilterMutex();
-  std::vector<AdapterPtr> &getAdapters();
+  std::vector<Adapter *> &getAdapters();
   ods_target_list &getOneapiDeviceSelectorTargets(const std::string &InitValue);
   XPTIRegistry &getXPTIRegistry();
   ThreadPool &getHostTaskThreadPool();
@@ -126,7 +125,7 @@ class GlobalHandler {
   InstWithLock<std::mutex> MPlatformToDefaultContextCacheMutex;
   InstWithLock<std::mutex> MPlatformMapMutex;
   InstWithLock<std::mutex> MFilterMutex;
-  InstWithLock<std::vector<AdapterPtr>> MAdapters;
+  InstWithLock<std::vector<Adapter *>> MAdapters;
   InstWithLock<ods_target_list> MOneapiDeviceSelectorTargets;
   InstWithLock<XPTIRegistry> MXPTIRegistry;
   // Thread pool for host task and event callbacks execution

sycl/source/detail/kernel_program_cache.hpp

@@ -112,25 +112,21 @@ class KernelProgramCache {
   };
 
   struct ProgramBuildResult : public BuildResult<ur_program_handle_t> {
-    std::weak_ptr<Adapter> AdapterWeakPtr;
-    ProgramBuildResult(const AdapterPtr &Adapter) : AdapterWeakPtr(Adapter) {
+    AdapterPtr MAdapter;
+    ProgramBuildResult(const AdapterPtr &AAdapter) : MAdapter(AAdapter) {
       Val = nullptr;
     }
-    ProgramBuildResult(const AdapterPtr &Adapter, BuildState InitialState)
-        : AdapterWeakPtr(Adapter) {
+    ProgramBuildResult(const AdapterPtr &AAdapter, BuildState InitialState)
+        : MAdapter(AAdapter) {
       Val = nullptr;
       this->State.store(InitialState);
     }
     ~ProgramBuildResult() {
       try {
         if (Val) {
-          AdapterPtr AdapterSharedPtr = AdapterWeakPtr.lock();
-          if (AdapterSharedPtr) {
-            ur_result_t Err =
-                AdapterSharedPtr->call_nocheck<UrApiKind::urProgramRelease>(
-                    Val);
-            __SYCL_CHECK_UR_CODE_NO_EXC(Err, AdapterSharedPtr->getBackend());
-          }
+          ur_result_t Err =
+              MAdapter->call_nocheck<UrApiKind::urProgramRelease>(Val);
+          __SYCL_CHECK_UR_CODE_NO_EXC(Err, MAdapter->getBackend());
         }
       } catch (std::exception &e) {
         __SYCL_REPORT_EXCEPTION_TO_STREAM("exception in ~ProgramBuildResult",
@@ -202,20 +198,16 @@ class KernelProgramCache {
   using KernelArgMaskPairT =
       std::pair<ur_kernel_handle_t, const KernelArgMask *>;
   struct KernelBuildResult : public BuildResult<KernelArgMaskPairT> {
-    std::weak_ptr<Adapter> AdapterWeakPtr;
-    KernelBuildResult(const AdapterPtr &Adapter) : AdapterWeakPtr(Adapter) {
+    AdapterPtr MAdapter;
+    KernelBuildResult(const AdapterPtr &AAdapter) : MAdapter(AAdapter) {
       Val.first = nullptr;
     }
     ~KernelBuildResult() {
       try {
         if (Val.first) {
-          AdapterPtr AdapterSharedPtr = AdapterWeakPtr.lock();
-          if (AdapterSharedPtr) {
-            ur_result_t Err =
-                AdapterSharedPtr->call_nocheck<UrApiKind::urKernelRelease>(
-                    Val.first);
-            __SYCL_CHECK_UR_CODE_NO_EXC(Err, AdapterSharedPtr->getBackend());
-          }
+          ur_result_t Err =
+              MAdapter->call_nocheck<UrApiKind::urKernelRelease>(Val.first);
+          __SYCL_CHECK_UR_CODE_NO_EXC(Err, MAdapter->getBackend());
         }
       } catch (std::exception &e) {
         __SYCL_REPORT_EXCEPTION_TO_STREAM("exception in ~KernelBuildResult", e);

sycl/source/detail/platform_impl.cpp

@@ -168,7 +168,7 @@ std::vector<platform> platform_impl::get_platforms() {
 
   // See which platform we want to be served by which adapter.
   // There should be just one adapter serving each backend.
-  std::vector<AdapterPtr> &Adapters = sycl::detail::ur::initializeUr();
+  std::vector<AdapterPtr> &Adapters = ur::initializeUr();
   std::vector<std::pair<platform, AdapterPtr>> PlatformsWithAdapter;
 
   // Then check backend-specific adapters
@@ -504,7 +504,7 @@ platform_impl::get_devices(info::device_type DeviceType) const {
     // analysis. Doing adjustment by simple copy of last device num from
     // previous platform.
     // Needs non const adapter reference.
-    std::vector<AdapterPtr> &Adapters = sycl::detail::ur::initializeUr();
+    std::vector<AdapterPtr> &Adapters = ur::initializeUr();
     auto It = std::find_if(Adapters.begin(), Adapters.end(),
                            [&Platform = MPlatform](AdapterPtr &Adapter) {
                              return Adapter->containsUrPlatform(Platform);

sycl/source/detail/platform_impl.hpp

@@ -39,8 +39,7 @@ class platform_impl : public std::enable_shared_from_this<platform_impl> {
   //
   // Platforms can only be created under `GlobalHandler`'s ownership via
   // `platform_impl::getOrMakePlatformImpl` method.
-  explicit platform_impl(ur_platform_handle_t APlatform,
-                         const std::shared_ptr<Adapter> &AAdapter)
+  explicit platform_impl(ur_platform_handle_t APlatform, Adapter *AAdapter)
       : MPlatform(APlatform), MAdapter(AAdapter) {
     // Find out backend of the platform
     ur_backend_t UrBackend = UR_BACKEND_UNKNOWN;

sycl/source/detail/program_manager/program_manager.cpp

@@ -1186,7 +1186,7 @@ FastKernelCacheValPtr ProgramManager::getOrCreateKernel(
     // nullptr for the mutex.
     auto [Kernel, ArgMask] = BuildF();
     return std::make_shared<FastKernelCacheVal>(
-        Kernel, nullptr, ArgMask, Program, *ContextImpl.getAdapter().get());
+        Kernel, nullptr, ArgMask, Program, *ContextImpl.getAdapter());
   }
 
   auto BuildResult = Cache.getOrBuild<errc::invalid>(GetCachedBuildF, BuildF);
@@ -1195,7 +1195,7 @@ FastKernelCacheValPtr ProgramManager::getOrCreateKernel(
   const KernelArgMaskPairT &KernelArgMaskPair = BuildResult->Val;
   auto ret_val = std::make_shared<FastKernelCacheVal>(
       KernelArgMaskPair.first, &(BuildResult->MBuildResultMutex),
-      KernelArgMaskPair.second, Program, *ContextImpl.getAdapter().get());
+      KernelArgMaskPair.second, Program, *ContextImpl.getAdapter());
   // If caching is enabled, one copy of the kernel handle will be
   // stored in FastKernelCacheVal, and one is in
   // KernelProgramCache::MKernelsPerProgramCache. To cover

sycl/source/detail/sycl_mem_obj_t.hpp

@@ -34,7 +34,7 @@ namespace detail {
 class context_impl;
 class event_impl;
 class Adapter;
-using AdapterPtr = std::shared_ptr<Adapter>;
+using AdapterPtr = Adapter *;
 
 using EventImplPtr = std::shared_ptr<event_impl>;
 

sycl/source/detail/ur.cpp

@@ -90,7 +90,7 @@ bool trace(TraceLevel Level) {
   return (TraceLevelMask & Level) == Level;
 }
 
-static void initializeAdapters(std::vector<AdapterPtr> &Adapters,
+static void initializeAdapters(std::vector<Adapter *> &Adapters,
                                ur_loader_config_handle_t LoaderConfig);
 
 bool XPTIInitDone = false;
@@ -117,7 +117,7 @@ std::vector<AdapterPtr> &initializeUr(ur_loader_config_handle_t LoaderConfig) {
   return GlobalHandler::instance().getAdapters();
 }
 
-static void initializeAdapters(std::vector<AdapterPtr> &Adapters,
+static void initializeAdapters(std::vector<Adapter *> &Adapters,
                                ur_loader_config_handle_t LoaderConfig) {
 #define CHECK_UR_SUCCESS(Call)                                                 \
   {                                                                            \
@@ -238,7 +238,7 @@ static void initializeAdapters(std::vector<AdapterPtr> &Adapters,
                                     sizeof(adapterBackend), &adapterBackend,
                                     nullptr));
     auto syclBackend = UrToSyclBackend(adapterBackend);
-    Adapters.emplace_back(std::make_shared<Adapter>(UrAdapter, syclBackend));
+    Adapters.emplace_back(new Adapter(UrAdapter, syclBackend));
 
     const char *env_value = std::getenv("UR_LOG_CALLBACK");
     if (env_value == nullptr || std::string(env_value) != "disabled") {
@@ -284,25 +284,25 @@ static void initializeAdapters(std::vector<AdapterPtr> &Adapters,
 }
 
 // Get the adapter serving given backend.
-template <backend BE> const AdapterPtr &getAdapter() {
-  static AdapterPtr *Adapter = nullptr;
-  if (Adapter)
-    return *Adapter;
+template <backend BE> AdapterPtr &getAdapter() {
+  static AdapterPtr adapterPtr = nullptr;
+  if (adapterPtr)
+    return adapterPtr;
 
-  std::vector<AdapterPtr> &Adapters = ur::initializeUr();
+  std::vector<AdapterPtr> Adapters = ur::initializeUr();
   for (auto &P : Adapters)
     if (P->hasBackend(BE)) {
-      Adapter = &P;
-      return *Adapter;
+      adapterPtr = P;
+      return adapterPtr;
     }
 
   throw exception(errc::runtime, "ur::getAdapter couldn't find adapter");
 }
 
-template const AdapterPtr &getAdapter<backend::opencl>();
-template const AdapterPtr &getAdapter<backend::ext_oneapi_level_zero>();
-template const AdapterPtr &getAdapter<backend::ext_oneapi_cuda>();
-template const AdapterPtr &getAdapter<backend::ext_oneapi_hip>();
+template AdapterPtr &getAdapter<backend::opencl>();
+template AdapterPtr &getAdapter<backend::ext_oneapi_level_zero>();
+template AdapterPtr &getAdapter<backend::ext_oneapi_cuda>();
+template AdapterPtr &getAdapter<backend::ext_oneapi_hip>();
 
 // Reads an integer value from ELF data.
 template <typename ResT>