wip

Author: martygrant

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

@@ -45,6 +45,7 @@ urContextCreate(uint32_t DeviceCount, const ur_device_handle_t *phDevices,
   try {
     ContextPtr = std::unique_ptr<ur_context_handle_t_>(
         new ur_context_handle_t_{phDevices, DeviceCount});
+    ContextPtr->RefCount.retain();
     *phContext = ContextPtr.release();
   } catch (ur_result_t Err) {
     return Err;
@@ -83,7 +84,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urContextGetInfo(
 
 UR_APIEXPORT ur_result_t UR_APICALL
 urContextRelease(ur_context_handle_t hContext) {
-  if (!hContext->RefCount.release()) {
+  if (hContext->RefCount.release()) {
     return UR_RESULT_SUCCESS;
   }
   hContext->invokeExtendedDeleters();

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

@@ -97,7 +97,7 @@ struct ur_context_handle_t_ : ur::cuda::handle_base {
   umf_memory_pool_handle_t MemoryPoolHost = nullptr;
 
   ur_context_handle_t_(const ur_device_handle_t *Devs, uint32_t NumDevices)
-      : handle_base(), Devices{Devs, Devs + NumDevices}, RefCount{1} {
+      : handle_base(), Devices{Devs, Devs + NumDevices}, RefCount(0) {
     // Create UMF CUDA memory provider for the host memory
     // (UMF_MEMORY_TYPE_HOST) from any device (Devices[0] is used here, because
     // it is guaranteed to exist).

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

@@ -256,7 +256,7 @@ struct ur_kernel_handle_t_ : ur::cuda::handle_base {
                       ur_context_handle_t Context)
       : handle_base(), Function{Func},
         FunctionWithOffsetParam{FuncWithOffsetParam}, Name{Name},
-        Context{Context}, Program{Program}, RefCount{1} {
+        Context{Context}, Program{Program} {
     urProgramRetain(Program);
     urContextRetain(Context);
 

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

@@ -345,17 +345,17 @@ struct ur_mem_handle_t_ : ur::cuda::handle_base {
   /// Constructs the UR mem handler for a non-typed allocation ("buffer")
   ur_mem_handle_t_(ur_context_handle_t Ctxt, ur_mem_flags_t MemFlags,
                    BufferMem::AllocMode Mode, void *HostPtr, size_t Size)
-      : handle_base(), Context{Ctxt}, RefCount{1}, MemFlags{MemFlags},
+      : handle_base(), Context{Ctxt}, MemFlags{MemFlags},
         HaveMigratedToDeviceSinceLastWrite(Context->Devices.size(), false),
         Mem{std::in_place_type<BufferMem>, Ctxt, this, Mode, HostPtr, Size} {
     urContextRetain(Context);
   };
 
   // Subbuffer constructor
   ur_mem_handle_t_(ur_mem_handle_t Parent, size_t SubBufferOffset)
-      : handle_base(), Context{Parent->Context}, RefCount{1},
-        MemFlags{Parent->MemFlags}, HaveMigratedToDeviceSinceLastWrite(
-                                        Parent->Context->Devices.size(), false),
+      : handle_base(), Context{Parent->Context}, MemFlags{Parent->MemFlags},
+        HaveMigratedToDeviceSinceLastWrite(Parent->Context->Devices.size(),
+                                           false),
         Mem{BufferMem{std::get<BufferMem>(Parent->Mem)}} {
     auto &SubBuffer = std::get<BufferMem>(Mem);
     SubBuffer.Parent = Parent;
@@ -376,7 +376,7 @@ struct ur_mem_handle_t_ : ur::cuda::handle_base {
   ur_mem_handle_t_(ur_context_handle_t Ctxt, ur_mem_flags_t MemFlags,
                    ur_image_format_t ImageFormat, ur_image_desc_t ImageDesc,
                    void *HostPtr)
-      : handle_base(), Context{Ctxt}, RefCount{1}, MemFlags{MemFlags},
+      : handle_base(), Context{Ctxt}, MemFlags{MemFlags},
         HaveMigratedToDeviceSinceLastWrite(Context->Devices.size(), false),
         Mem{std::in_place_type<SurfaceMem>,
             Ctxt,

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

@@ -50,9 +50,9 @@ struct ur_program_handle_t_ : ur::cuda::handle_base {
 
   ur_program_handle_t_(ur_context_handle_t Context, ur_device_handle_t Device)
       : handle_base(), Module{nullptr}, Binary{}, BinarySizeInBytes{0},
-        RefCount{1}, Context{Context}, Device{Device},
-        KernelReqdWorkGroupSizeMD{}, KernelMaxWorkGroupSizeMD{},
-        KernelMaxLinearWorkGroupSizeMD{}, KernelReqdSubGroupSizeMD{} {
+        Context{Context}, Device{Device}, KernelReqdWorkGroupSizeMD{},
+        KernelMaxWorkGroupSizeMD{}, KernelMaxLinearWorkGroupSizeMD{},
+        KernelReqdSubGroupSizeMD{} {
     urContextRetain(Context);
 
     // When the log is queried we use strnlen(InfoLog), so it needs to be

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

@@ -34,7 +34,7 @@ struct ur_sampler_handle_t_ : ur::cuda::handle_base {
   ur_context_handle_t Context;
 
   ur_sampler_handle_t_(ur_context_handle_t Context)
-      : handle_base(), RefCount(1), Props(0), MinMipmapLevelClamp(0.0f),
+      : handle_base(), Props(0), MinMipmapLevelClamp(0.0f),
         MaxMipmapLevelClamp(0.0f), MaxAnisotropy(0.0f), Context(Context) {}
 
   ur_bool_t isNormalizedCoords() const noexcept {