Use function-local static variable instead.

microsoft · Jan 7, 2022 · 0a68c6a · 0a68c6a
1 parent c73f198
commit 0a68c6a
Show file tree

Hide file tree

Showing 4 changed files with 33 additions and 73 deletions.
diff --git a/onnxruntime/core/common/shared_ptr_thread_safe_wrapper.h b/onnxruntime/core/common/shared_ptr_thread_safe_wrapper.h
diff --git a/onnxruntime/core/providers/cuda/cuda_execution_provider.cc b/onnxruntime/core/providers/cuda/cuda_execution_provider.cc
@@ -3,7 +3,6 @@
 
 #include "core/providers/shared_library/provider_api.h"
 #include "core/providers/cuda/cuda_execution_provider.h"
-#include "core/common/shared_ptr_thread_safe_wrapper.h"
 #include "core/providers/cuda/cuda_common.h"
 #include "core/providers/cuda/cuda_allocator.h"
 #include "core/providers/cuda/cuda_fence.h"
@@ -2074,20 +2073,23 @@ static Status RegisterCudaKernels(KernelRegistry& kernel_registry) {
 
 }  // namespace cuda
 
-static std::shared_ptr<KernelRegistry> CreateCudaKernelRegistry() {
-  std::shared_ptr<KernelRegistry> registry = KernelRegistry::Create();
-  ORT_THROW_IF_ERROR(cuda::RegisterCudaKernels(*registry));
-  return registry;
-}
+static std::shared_ptr<KernelRegistry>& CudaKernelRegistry() {
+  // static local variable ensures thread-safe initialization
+  static std::shared_ptr<KernelRegistry> cuda_kernel_registry = []() {
+    std::shared_ptr<KernelRegistry> registry = KernelRegistry::Create();
+    ORT_THROW_IF_ERROR(cuda::RegisterCudaKernels(*registry));
+    return registry;
+  }();
 
-static SharedPtrThreadSafeWrapper<KernelRegistry> s_kernel_registry{&CreateCudaKernelRegistry};
+  return cuda_kernel_registry;
+}
 
 void Shutdown_DeleteRegistry() {
-  s_kernel_registry.Reset();
+  CudaKernelRegistry().reset();
 }
 
 std::shared_ptr<KernelRegistry> CUDAExecutionProvider::GetKernelRegistry() const {
-  return s_kernel_registry.GetInitialized();
+  return CudaKernelRegistry();
 }
 
 static bool RNNNeedFallbackToCPU(const onnxruntime::Node& node,

diff --git a/onnxruntime/core/providers/rocm/rocm_execution_provider.cc b/onnxruntime/core/providers/rocm/rocm_execution_provider.cc
@@ -3,7 +3,6 @@
 
 #include "core/providers/shared_library/provider_api.h"
 #include "core/providers/rocm/rocm_execution_provider.h"
-#include "core/common/shared_ptr_thread_safe_wrapper.h"
 #include "core/providers/rocm/rocm_common.h"
 #include "core/providers/rocm/rocm_allocator.h"
 #include "core/providers/rocm/rocm_fence.h"
@@ -2067,20 +2066,23 @@ static Status RegisterRocmKernels(KernelRegistry& kernel_registry) {
 
 }  // namespace rocm
 
-static std::shared_ptr<KernelRegistry> CreateRocmKernelRegistry() {
-  std::shared_ptr<KernelRegistry> registry = KernelRegistry::Create();
-  ORT_THROW_IF_ERROR(rocm::RegisterRocmKernels(*registry));
-  return registry;
-}
+static std::shared_ptr<KernelRegistry>& RocmKernelRegistry() {
+  // static local variable ensures thread-safe initialization
+  static std::shared_ptr<KernelRegistry> rocm_kernel_registry = []() {
+    std::shared_ptr<KernelRegistry> registry = KernelRegistry::Create();
+    ORT_THROW_IF_ERROR(rocm::RegisterRocmKernels(*registry));
+    return registry;
+  }();
 
-static SharedPtrThreadSafeWrapper<KernelRegistry> s_kernel_registry{&CreateRocmKernelRegistry};
+  return rocm_kernel_registry;
+}
 
 void Shutdown_DeleteRegistry() {
-  s_kernel_registry.Reset();
+  RocmKernelRegistry().reset();
 }
 
 std::shared_ptr<KernelRegistry> ROCMExecutionProvider::GetKernelRegistry() const {
-  return s_kernel_registry.GetInitialized();
+  return RocmKernelRegistry();
 }
 
 static bool CastNeedFallbackToCPU(const onnxruntime::Node& node) {

diff --git a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
@@ -7,7 +7,6 @@
 #define ORT_API_MANUAL_INIT
 #include "core/session/onnxruntime_cxx_api.h"
 #include "core/common/safeint.h"
-#include "core/common/shared_ptr_thread_safe_wrapper.h"
 #include "tensorrt_execution_provider.h"
 #include "core/providers/cuda/shared_inc/cuda_call.h"
 #include "core/providers/cuda/math/unary_elementwise_ops_impl.h"
@@ -377,20 +376,23 @@ static Status RegisterTensorrtKernels(KernelRegistry& kernel_registry) {
   return Status::OK();
 }
 
-static std::shared_ptr<KernelRegistry> CreateTensorrtKernelRegistry() {
-  std::shared_ptr<KernelRegistry> registry = KernelRegistry::Create();
-  ORT_THROW_IF_ERROR(RegisterTensorrtKernels(*registry));
-  return registry;
-}
+static std::shared_ptr<KernelRegistry>& TensorrtKernelRegistry() {
+  // static local variable ensures thread-safe initialization
+  static std::shared_ptr<KernelRegistry> tensorrt_kernel_registry = []() {
+    std::shared_ptr<KernelRegistry> registry = KernelRegistry::Create();
+    ORT_THROW_IF_ERROR(RegisterTensorrtKernels(*registry));
+    return registry;
+  }();
 
-static SharedPtrThreadSafeWrapper<KernelRegistry> s_kernel_registry{&CreateTensorrtKernelRegistry};
+  return tensorrt_kernel_registry;
+}
 
 void Shutdown_DeleteRegistry() {
-  s_kernel_registry.Reset();
+  TensorrtKernelRegistry().reset();
 }
 
 std::shared_ptr<KernelRegistry> TensorrtExecutionProvider::GetKernelRegistry() const {
-  return s_kernel_registry.GetInitialized();
+  return TensorrtKernelRegistry();
 }
 
 // Per TensorRT documentation, logger needs to be a singleton.