[UR] Remove unnecessary and confusing unique_ptr usage (#17144)

`std::unique_ptr` was being used as a scoped deleter for a UMF value,
but this was both confusing and more expensive than it needed to be.

Here we introduce the simple `OnScopeExit` class which calls a callable
when it goes out of scope, and use it here to simplify the meaning and
overhead of the CUDA adapter's UMF pool allocation.

Author: ldrumm

unified-runtime/source/adapters/cuda/common.hpp

@@ -73,10 +73,6 @@ void assertion(bool Condition, const char *Message = nullptr);
 
 namespace umf {
 
-using cuda_params_unique_handle_t = std::unique_ptr<
-    umf_cuda_memory_provider_params_t,
-    std::function<umf_result_t(umf_cuda_memory_provider_params_handle_t)>>;
-
 inline umf_result_t setCUMemoryProviderParams(
     umf_cuda_memory_provider_params_handle_t CUMemoryProviderParams,
     int cuDevice, void *cuContext, umf_usm_memory_type_t memType) {

unified-runtime/source/adapters/cuda/context.hpp

@@ -80,28 +80,25 @@ static ur_result_t
 CreateHostMemoryProviderPool(ur_device_handle_t_ *DeviceHandle,
                              umf_memory_provider_handle_t *MemoryProviderHost,
                              umf_memory_pool_handle_t *MemoryPoolHost) {
-  umf_cuda_memory_provider_params_handle_t CUMemoryProviderParams = nullptr;
 
   *MemoryProviderHost = nullptr;
   CUcontext context = DeviceHandle->getNativeContext();
 
+  umf_cuda_memory_provider_params_handle_t CUMemoryProviderParams = nullptr;
   umf_result_t UmfResult =
       umfCUDAMemoryProviderParamsCreate(&CUMemoryProviderParams);
   UMF_RETURN_UR_ERROR(UmfResult);
+  OnScopeExit Cleanup(
+      [=]() { umfCUDAMemoryProviderParamsDestroy(CUMemoryProviderParams); });
 
-  umf::cuda_params_unique_handle_t CUMemoryProviderParamsUnique(
-      CUMemoryProviderParams, umfCUDAMemoryProviderParamsDestroy);
-
-  UmfResult = umf::setCUMemoryProviderParams(CUMemoryProviderParamsUnique.get(),
-                                             0 /* cuDevice */, context,
-                                             UMF_MEMORY_TYPE_HOST);
+  UmfResult = umf::setCUMemoryProviderParams(
+      CUMemoryProviderParams, 0 /* cuDevice */, context, UMF_MEMORY_TYPE_HOST);
   UMF_RETURN_UR_ERROR(UmfResult);
 
   // create UMF CUDA memory provider and pool for the host memory
   // (UMF_MEMORY_TYPE_HOST)
-  UmfResult = umfMemoryProviderCreate(umfCUDAMemoryProviderOps(),
-                                      CUMemoryProviderParamsUnique.get(),
-                                      MemoryProviderHost);
+  UmfResult = umfMemoryProviderCreate(
+      umfCUDAMemoryProviderOps(), CUMemoryProviderParams, MemoryProviderHost);
   UMF_RETURN_UR_ERROR(UmfResult);
 
   UmfResult = umfPoolCreate(umfProxyPoolOps(), *MemoryProviderHost, nullptr, 0,

unified-runtime/source/adapters/cuda/platform.cpp

@@ -27,8 +27,8 @@ CreateDeviceMemoryProvidersPools(ur_platform_handle_t_ *Platform) {
       umfCUDAMemoryProviderParamsCreate(&CUMemoryProviderParams);
   UMF_RETURN_UR_ERROR(UmfResult);
 
-  umf::cuda_params_unique_handle_t CUMemoryProviderParamsUnique(
-      CUMemoryProviderParams, umfCUDAMemoryProviderParamsDestroy);
+  OnScopeExit Cleanup(
+      [=]() { umfCUDAMemoryProviderParamsDestroy(CUMemoryProviderParams); });
 
   for (auto &Device : Platform->Devices) {
     ur_device_handle_t_ *device_handle = Device.get();
@@ -37,25 +37,23 @@ CreateDeviceMemoryProvidersPools(ur_platform_handle_t_ *Platform) {
 
     // create UMF CUDA memory provider for the device memory
     // (UMF_MEMORY_TYPE_DEVICE)
-    UmfResult =
-        umf::setCUMemoryProviderParams(CUMemoryProviderParamsUnique.get(),
-                                       device, context, UMF_MEMORY_TYPE_DEVICE);
+    UmfResult = umf::setCUMemoryProviderParams(CUMemoryProviderParams, device,
+                                               context, UMF_MEMORY_TYPE_DEVICE);
     UMF_RETURN_UR_ERROR(UmfResult);
 
     UmfResult = umfMemoryProviderCreate(umfCUDAMemoryProviderOps(),
-                                        CUMemoryProviderParamsUnique.get(),
+                                        CUMemoryProviderParams,
                                         &device_handle->MemoryProviderDevice);
     UMF_RETURN_UR_ERROR(UmfResult);
 
     // create UMF CUDA memory provider for the shared memory
     // (UMF_MEMORY_TYPE_SHARED)
-    UmfResult =
-        umf::setCUMemoryProviderParams(CUMemoryProviderParamsUnique.get(),
-                                       device, context, UMF_MEMORY_TYPE_SHARED);
+    UmfResult = umf::setCUMemoryProviderParams(CUMemoryProviderParams, device,
+                                               context, UMF_MEMORY_TYPE_SHARED);
     UMF_RETURN_UR_ERROR(UmfResult);
 
     UmfResult = umfMemoryProviderCreate(umfCUDAMemoryProviderOps(),
-                                        CUMemoryProviderParamsUnique.get(),
+                                        CUMemoryProviderParams,
                                         &device_handle->MemoryProviderShared);
     UMF_RETURN_UR_ERROR(UmfResult);
 

unified-runtime/source/common/ur_util.hpp

@@ -25,6 +25,7 @@
 #include <sstream>
 #include <string.h>
 #include <string>
+#include <type_traits>
 #include <vector>
 
 int ur_getpid(void);
@@ -484,6 +485,27 @@ template <typename T> class AtomicSingleton {
   }
 };
 
+// A type that will call your function on exit of the current scope. Useful for
+// avoiding leaks with C APIs.
+// e.g.
+//
+// {
+//    OnScopeExit _([]() { puts("out of scope"); };
+// }
+//
+// Decent compilers (and MSVC) will optimize the above to a simple call to puts
+template <class Callable> struct OnScopeExit {
+  const Callable Fn;
+  OnScopeExit(Callable Fn) : Fn(Fn) {}
+  // We don't want copies because this should only ever call the user function
+  // once and we don't know where we might end up if we allow copies
+  OnScopeExit(OnScopeExit &) = delete;
+  OnScopeExit(OnScopeExit &&) = delete;
+  ~OnScopeExit() { Fn(); }
+};
+// Silence a warning about unintended deduction
+template <class Callable> OnScopeExit(Callable) -> OnScopeExit<Callable>;
+
 template <typename Numeric>
 static inline std::string groupDigits(Numeric numeric) {
   auto number = std::to_string(numeric);