From f20258e9ed584ac3b3e39fbf2c8e72ec98d9840e Mon Sep 17 00:00:00 2001 From: Changming Sun Date: Wed, 6 Feb 2019 14:31:02 -0800 Subject: [PATCH] Delete dead code --- .../core/optimizer/insert_cast_transformer.h | 12 ------------ onnxruntime/core/session/inference_session.cc | 4 +--- .../test/framework/insert_cast_transformer_test.cc | 14 -------------- 3 files changed, 1 insertion(+), 29 deletions(-) diff --git a/onnxruntime/core/optimizer/insert_cast_transformer.h b/onnxruntime/core/optimizer/insert_cast_transformer.h index a8d27223da..24668f852b 100644 --- a/onnxruntime/core/optimizer/insert_cast_transformer.h +++ b/onnxruntime/core/optimizer/insert_cast_transformer.h @@ -15,23 +15,11 @@ class InsertCastTransformer : public onnxruntime::GraphTransformer { force_cpu_fp32_(true) { } - void AddKernelRegistries(const std::vector& kernels) { - for (auto* kernel : kernels) { - if (kernel) - kernels_registries_.push_back(kernel); - } - } - - void AddKernelRegistry(const KernelRegistry& kernel) { - kernels_registries_.push_back(&kernel); - } - private: Status ApplyImpl(onnxruntime::Graph& graph, bool& modified, int graph_level) const override; bool NeedInsertCast(const onnxruntime::Node* node, const onnxruntime::NodeArg* input) const; - std::vector kernels_registries_; // Currently because we only have very few cpu kernels support float16, place those nodes on float16 // will introduce many cast between fp32 and fp16, which will slow the execution. // A better solution is to have a cost model to evaluate does it works to place the node on float16. diff --git a/onnxruntime/core/session/inference_session.cc b/onnxruntime/core/session/inference_session.cc index 93d98b8aad..c98218ab05 100644 --- a/onnxruntime/core/session/inference_session.cc +++ b/onnxruntime/core/session/inference_session.cc @@ -341,9 +341,7 @@ class InferenceSession::Impl { // The 1st ones should have already been registered via session-level API into KernelRegistryManager. // // Register 2nd registries into KernelRegistryManager. - kernel_registry_manager_.RegisterKernels(execution_providers_); - - insert_cast_transformer_.AddKernelRegistries(kernel_registry_manager_.GetAllKernelRegistries()); + kernel_registry_manager_.RegisterKernels(execution_providers_, KernelRegistryPriority::LowPriority); SessionStateInitializer session_initializer{graph, session_state_, execution_providers_, kernel_registry_manager_}; diff --git a/onnxruntime/test/framework/insert_cast_transformer_test.cc b/onnxruntime/test/framework/insert_cast_transformer_test.cc index a4a0ffebdf..79ddf5f1b3 100644 --- a/onnxruntime/test/framework/insert_cast_transformer_test.cc +++ b/onnxruntime/test/framework/insert_cast_transformer_test.cc @@ -32,14 +32,7 @@ TEST(TransformerTest, InsertCastGPUTest) { auto status = graph.Resolve(); ASSERT_TRUE(status.IsOK()) << status.ErrorMessage(); - auto cpu_execution_provider = TestCPUExecutionProvider(); InsertCastTransformer transformer("Test"); - transformer.AddKernelRegistry(*cpu_execution_provider->GetKernelRegistry().get()); - -#ifdef USE_CUDA - auto cuda_execution_provider = TestCudaExecutionProvider(); - transformer.AddKernelRegistry(*cuda_execution_provider->GetKernelRegistry().get()); -#endif bool modified = true; status = transformer.Apply(graph, modified); @@ -87,14 +80,7 @@ TEST(TransformerTest, InsertCastAllCPUTest) { auto status = graph.Resolve(); ASSERT_TRUE(status.IsOK()) << status.ErrorMessage(); - auto cpu_execution_provider = TestCPUExecutionProvider(); InsertCastTransformer transformer("Test"); - transformer.AddKernelRegistry(*cpu_execution_provider->GetKernelRegistry().get()); - -#ifdef USE_CUDA - auto cuda_execution_provider = TestCudaExecutionProvider(); - transformer.AddKernelRegistry(*cuda_execution_provider->GetKernelRegistry().get()); -#endif bool modified = true; EXPECT_TRUE(transformer.Apply(graph, modified).IsOK());