[UR][CUDA][HIP] Remove unnecessary device retain/release (#17937)

`urDeviceRetain` and `urDeviceRelease` are no-ops, so this patch removes
all uses.

Uses of it in the context will be removed in
#17571

Author: npmiller

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

@@ -61,17 +61,13 @@ ur_exp_command_buffer_handle_t_::ur_exp_command_buffer_handle_t_(
       CudaGraph{nullptr}, CudaGraphExec{nullptr}, RefCount{1},
       NextSyncPoint{0} {
   urContextRetain(Context);
-  urDeviceRetain(Device);
 }
 
 /// The ur_exp_command_buffer_handle_t_ destructor releases
 /// all the memory objects allocated for command_buffer managment
 ur_exp_command_buffer_handle_t_::~ur_exp_command_buffer_handle_t_() {
   // Release the memory allocated to the Context stored in the command_buffer
   UR_CALL_NOCHECK(urContextRelease(Context));
-
-  // Release the device
-  UR_CALL_NOCHECK(urDeviceRelease(Device));
 }
 
 // This may throw so it must be called from within a try...catch

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

@@ -1198,8 +1198,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
 
 /// \return PI_SUCCESS if the function is executed successfully
 /// CUDA devices are always root devices so retain always returns success.
-UR_APIEXPORT ur_result_t UR_APICALL urDeviceRetain(ur_device_handle_t hDevice) {
-  std::ignore = hDevice;
+UR_APIEXPORT ur_result_t UR_APICALL urDeviceRetain(ur_device_handle_t) {
   return UR_RESULT_SUCCESS;
 }
 
@@ -1211,9 +1210,7 @@ urDevicePartition(ur_device_handle_t, const ur_device_partition_properties_t *,
 
 /// \return UR_RESULT_SUCCESS always since CUDA devices are always root
 /// devices.
-UR_APIEXPORT ur_result_t UR_APICALL
-urDeviceRelease(ur_device_handle_t hDevice) {
-  std::ignore = hDevice;
+UR_APIEXPORT ur_result_t UR_APICALL urDeviceRelease(ur_device_handle_t) {
   return UR_RESULT_SUCCESS;
 }
 

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

@@ -36,13 +36,9 @@ struct ur_physical_mem_handle_t_ {
       : RefCount(1), PhysicalMem(PhysMem), Context(Ctx), Device(Device),
         Size(Size), Properties(Properties) {
     urContextRetain(Context);
-    urDeviceRetain(Device);
   }
 
-  ~ur_physical_mem_handle_t_() {
-    urContextRelease(Context);
-    urDeviceRelease(Device);
-  }
+  ~ur_physical_mem_handle_t_() { urContextRelease(Context); }
 
   native_type get() const noexcept { return PhysicalMem; }
 

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

@@ -53,13 +53,9 @@ struct ur_program_handle_t_ {
         KernelMaxWorkGroupSizeMD{}, KernelMaxLinearWorkGroupSizeMD{},
         KernelReqdSubGroupSizeMD{} {
     urContextRetain(Context);
-    urDeviceRetain(Device);
   }
 
-  ~ur_program_handle_t_() {
-    urContextRelease(Context);
-    urDeviceRelease(Device);
-  }
+  ~ur_program_handle_t_() { urContextRelease(Context); }
 
   ur_result_t setMetadata(const ur_program_metadata_t *Metadata, size_t Length);
 

unified-runtime/source/adapters/hip/command_buffer.cpp

@@ -26,7 +26,6 @@ ur_exp_command_buffer_handle_t_::ur_exp_command_buffer_handle_t_(
     : Context(hContext), Device(hDevice), IsUpdatable(IsUpdatable),
       HIPGraph{nullptr}, HIPGraphExec{nullptr}, RefCount{1}, NextSyncPoint{0} {
   urContextRetain(hContext);
-  urDeviceRetain(hDevice);
 }
 
 /// The ur_exp_command_buffer_handle_t_ destructor releases
@@ -35,9 +34,6 @@ ur_exp_command_buffer_handle_t_::~ur_exp_command_buffer_handle_t_() {
   // Release the memory allocated to the Context stored in the command_buffer
   UR_CALL_NOCHECK(urContextRelease(Context));
 
-  // Release the device
-  UR_CALL_NOCHECK(urDeviceRelease(Device));
-
   // Release the memory allocated to the HIPGraph
   (void)hipGraphDestroy(HIPGraph);
 

unified-runtime/source/adapters/hip/program.hpp

@@ -52,13 +52,9 @@ struct ur_program_handle_t_ {
         Context{Ctxt}, Device{Device}, KernelReqdWorkGroupSizeMD{},
         KernelReqdSubGroupSizeMD{} {
     urContextRetain(Context);
-    urDeviceRetain(Device);
   }
 
-  ~ur_program_handle_t_() {
-    urContextRelease(Context);
-    urDeviceRelease(Device);
-  }
+  ~ur_program_handle_t_() { urContextRelease(Context); }
 
   ur_result_t setMetadata(const ur_program_metadata_t *Metadata, size_t Length);