fix some calls to release objects

Author: omarahmed1111

source/adapters/level_zero/adapter_lib_init_windows.cpp

@@ -13,6 +13,25 @@
 
 #include <windows.h>
 
+// On windows, UR is destructed before sycl-rt have destructed the objects, so
+// we need to clear the leftover memory before UR destruction.
+ur_result_t deleteCacheOnDestruction() {
+  std::lock_guard<std::mutex> Lock{GlobalAdapter->Mutex};
+  const auto *platforms = GlobalAdapter->PlatformCache->get_value();
+  for (const auto &p : *platforms) {
+    std::scoped_lock<ur_shared_mutex> ContextsLock(p->ContextsMutex);
+    while (!p->Contexts.empty()) {
+      ur_context_handle_t &ctx = p->Contexts.front();
+      ctx->deleteCachedObjectsOnDestruction();
+      uint32_t RefCount = ctx->RefCount.load();
+      while (RefCount--) {
+        UR_CALL(urContextRelease(ctx));
+      }
+      deleteFromCachedList<ur_context_handle_t>(ctx, p->Contexts);
+    }
+  }
+}
+
 BOOL WINAPI DllMain(HINSTANCE hinstDLL, // handle to DLL module
                     DWORD fdwReason,    // reason for calling function
                     LPVOID lpReserved)  // reserved
@@ -21,18 +40,7 @@ BOOL WINAPI DllMain(HINSTANCE hinstDLL, // handle to DLL module
   case DLL_PROCESS_ATTACH:
     break;
   case DLL_PROCESS_DETACH: {
-    std::lock_guard<std::mutex> Lock{GlobalAdapter->Mutex};
-    const auto *platforms = GlobalAdapter->PlatformCache->get_value();
-    for (const auto &p : *platforms) {
-      std::scoped_lock<ur_shared_mutex> ContextsLock(p->ContextsMutex);
-      while (!p->Contexts.empty()) {
-        ur_context_handle_t &ctx = p->Contexts.front();
-        ctx->deleteCachedObjectsOnDestruction();
-        UR_CALL(urContextRelease(ctx));
-        deleteFromCachedList<ur_context_handle_t>(ctx, p->Contexts);
-      }
-    }
-    break;
+    return deleteCacheOnDestruction() == UR_RESULT_SUCCESS;
   }
   case DLL_THREAD_ATTACH:
     break;

source/adapters/level_zero/context.cpp

@@ -368,13 +368,12 @@ ur_result_t ContextReleaseHelper(ur_context_handle_t Context) {
       Contexts.erase(It);
   };
 
-#ifdef _WIN32
   DeleteFromContextsCache();
-#else
+
   if (IndirectAccessTrackingEnabled) {
     DeleteFromContextsCache();
   }
-#endif
+
   ze_context_handle_t DestroyZeContext =
       Context->OwnNativeHandle ? Context->ZeContext : nullptr;
 
@@ -857,3 +856,14 @@ ur_context_handle_t_::getDevices() const {
 ze_context_handle_t ur_context_handle_t_::getZeHandle() const {
   return ZeContext;
 }
+
+void ur_context_handle_t_::deleteCachedObjectsOnDestruction() {
+  while (!KernelsCache.empty()) {
+    ur_kernel_handle_t &kernel = KernelsCache.front();
+    uint32_t RefCount = kernel->RefCount.load();
+    while (RefCount--) {
+      UR_CALL_THROWS(urKernelRelease(kernel));
+    }
+    deleteFromCachedList(kernel, KernelsCache);
+  }
+}

source/adapters/level_zero/context.hpp

@@ -23,6 +23,7 @@
 #include <zes_api.h>
 
 #include "common.hpp"
+#include "kernel.hpp"
 #include "queue.hpp"
 
 #include <umf_helpers.hpp>
@@ -311,13 +312,7 @@ struct ur_context_handle_t_ : _ur_object {
   // Get handle to the L0 context
   ze_context_handle_t getZeHandle() const;
 
-  void deleteCachedObjectsOnDestruction() {
-    while (!KernelsCache.empty()) {
-      ur_kernel_handle_t &kernel = KernelsCache.front();
-      UR_CALL_THROWS(urKernelRelease(kernel));
-      deleteFromCachedList(kernel, KernelsCache);
-    }
-  }
+  void deleteCachedObjectsOnDestruction();
 
 private:
   // Get the cache of events for a provided scope and profiling mode.