[UR][L0] Init/deinit L0 adapter on first/last use (#19089)

Instead of using a global constructor to initialize the L0 adapter,
do it in the first call to `urAdapterGet`. Likewise, instead of
de-initing it as a global destructor, do it in the last call to
`urAdapterRelease`.

As well as not doing L0 initialization where the user is not using
L0, it also allows `urAdapterRelease` to be called in a global
destructor (e.g. what the SYCL runtime does) without worrying
about global destructor order.

Author: RossBrunton

unified-runtime/source/adapters/level_zero/CMakeLists.txt

@@ -90,13 +90,6 @@ if(UR_BUILD_ADAPTER_L0)
         endforeach()
     endif()
 
-    if(NOT WIN32)
-        target_sources(ur_adapter_level_zero
-            PRIVATE
-                ${CMAKE_CURRENT_SOURCE_DIR}/adapter_lib_init_linux.cpp
-        )
-    endif()
-
     # TODO: fix level_zero adapter conversion warnings
     target_compile_options(ur_adapter_level_zero PRIVATE
         $<$<CXX_COMPILER_ID:MSVC>:/wd4805 /wd4244>
@@ -195,11 +188,6 @@ if(UR_BUILD_ADAPTER_L0_V2)
         # api.cpp contains NOT_SUPPORTED functions-only
         set_source_files_properties(${CMAKE_CURRENT_SOURCE_DIR}/v2/api.cpp
             PROPERTIES APPEND_STRING PROPERTY COMPILE_FLAGS "-Wno-unused-parameter")
-
-        target_sources(ur_adapter_level_zero_v2
-            PRIVATE
-                ${CMAKE_CURRENT_SOURCE_DIR}/adapter_lib_init_linux.cpp
-        )
     endif()
 
     target_compile_definitions(ur_adapter_level_zero_v2 PUBLIC UR_STATIC_LEVEL_ZERO)

unified-runtime/source/adapters/level_zero/adapter.cpp

@@ -665,24 +665,19 @@ ur_result_t urAdapterGet(
     ur_adapter_handle_t *Adapters,
     /// [out][optional] returns the total number of adapters available.
     uint32_t *NumAdapters) {
+  static std::mutex AdapterConstructionMutex{};
+
   if (NumEntries > 0 && Adapters) {
-    if (GlobalAdapter) {
-      std::lock_guard<std::mutex> Lock{GlobalAdapter->Mutex};
-      if (GlobalAdapter->RefCount++ == 0) {
-        adapterStateInit();
-      }
-    } else {
-      // If the GetAdapter is called after the Library began or was torndown,
-      // then temporarily create a new Adapter handle and register a new
-      // cleanup.
+    std::lock_guard<std::mutex> Lock{AdapterConstructionMutex};
+
+    if (!GlobalAdapter) {
       GlobalAdapter = new ur_adapter_handle_t_();
-      std::lock_guard<std::mutex> Lock{GlobalAdapter->Mutex};
-      if (GlobalAdapter->RefCount++ == 0) {
-        adapterStateInit();
-      }
-      std::atexit(globalAdapterOnDemandCleanup);
     }
     *Adapters = GlobalAdapter;
+
+    if (GlobalAdapter->RefCount++ == 0) {
+      adapterStateInit();
+    }
   }
 
   if (NumAdapters) {
@@ -692,29 +687,31 @@ ur_result_t urAdapterGet(
   return UR_RESULT_SUCCESS;
 }
 
-ur_result_t urAdapterRelease(ur_adapter_handle_t) {
-  // Check first if the Adapter pointer is valid
-  if (GlobalAdapter) {
-    std::lock_guard<std::mutex> Lock{GlobalAdapter->Mutex};
-    if (--GlobalAdapter->RefCount == 0) {
-      auto result = adapterStateTeardown();
+ur_result_t urAdapterRelease([[maybe_unused]] ur_adapter_handle_t Adapter) {
+  assert(GlobalAdapter && GlobalAdapter == Adapter);
+
+  // NOTE: This does not require guarding with a mutex; the instant the ref
+  // count hits zero, both Get and Retain are UB.
+  if (--GlobalAdapter->RefCount == 0) {
+    auto result = adapterStateTeardown();
 #ifdef UR_STATIC_LEVEL_ZERO
-      // Given static linking of the L0 Loader, we must delay the loader's
-      // destruction of its context until after the UR Adapter is destroyed.
-      zelLoaderContextTeardown();
+    // Given static linking of the L0 Loader, we must delay the loader's
+    // destruction of its context until after the UR Adapter is destroyed.
+    zelLoaderContextTeardown();
 #endif
-      return result;
-    }
+
+    delete GlobalAdapter;
+    GlobalAdapter = nullptr;
+
+    return result;
   }
 
   return UR_RESULT_SUCCESS;
 }
 
-ur_result_t urAdapterRetain(ur_adapter_handle_t) {
-  if (GlobalAdapter) {
-    std::lock_guard<std::mutex> Lock{GlobalAdapter->Mutex};
-    GlobalAdapter->RefCount++;
-  }
+ur_result_t urAdapterRetain([[maybe_unused]] ur_adapter_handle_t Adapter) {
+  assert(GlobalAdapter && GlobalAdapter == Adapter);
+  GlobalAdapter->RefCount++;
 
   return UR_RESULT_SUCCESS;
 }

unified-runtime/source/adapters/level_zero/adapter.hpp

@@ -14,7 +14,6 @@
 #include <atomic>
 #include <loader/ur_loader.hpp>
 #include <loader/ze_loader.h>
-#include <mutex>
 #include <optional>
 #include <ur/ur.hpp>
 #include <ze_api.h>
@@ -28,7 +27,6 @@ class ur_legacy_sink;
 struct ur_adapter_handle_t_ : ur::handle_base<ur::level_zero::ddi_getter> {
   ur_adapter_handle_t_();
   std::atomic<uint32_t> RefCount = 0;
-  std::mutex Mutex;
 
   zes_pfnDriverGetDeviceByUuidExp_t getDeviceByUUIdFunctionPtr = nullptr;
   zes_pfnDriverGet_t getSysManDriversFunctionPtr = nullptr;