From ade4fa108f2061734d7e9fd6f48f5e1991b46bc2 Mon Sep 17 00:00:00 2001
From: KeDengMS <kedeng@microsoft.com>
Date: Thu, 5 Mar 2020 14:40:22 -0800
Subject: [PATCH] Disable delayload for cuda dlls (#3147)

This change fixes #3129. When running onnxruntime as dll on Windows, CUDA does some internal cleanups when process exits. After this, any call to CUDA would cause crash. Delayload makes thread_local destructor to happen after CUDA cleanup, thus the crash.
---
 cmake/CMakeLists.txt                              | 13 +++++++------
 onnxruntime/test/shared_lib/test_model_loading.cc |  9 +++++++++
 2 files changed, 16 insertions(+), 6 deletions(-)

diff --git a/cmake/CMakeLists.txt b/cmake/CMakeLists.txt
index cd586f22c6..c0ccab69ee 100644
--- a/cmake/CMakeLists.txt
+++ b/cmake/CMakeLists.txt
@@ -724,12 +724,13 @@ if (onnxruntime_USE_CUDA)
   if (WIN32)
     link_directories(${onnxruntime_CUDNN_HOME}/lib/x64)
 
-    file(GLOB cuda_dll_paths "${onnxruntime_CUDA_HOME}/bin/cublas64_*" "${onnxruntime_CUDA_HOME}/bin/cudart64_*" "${onnxruntime_CUDA_HOME}/bin/curand64_*")
-    set(onnxruntime_DELAYLOAD_FLAGS "${onnxruntime_DELAYLOAD_FLAGS} /DELAYLOAD:cudnn64_7.dll")
-    foreach(cuda_dll_path ${cuda_dll_paths})
-        get_filename_component(cuda_dll_file_name ${cuda_dll_path} NAME)
-        set(onnxruntime_DELAYLOAD_FLAGS "${onnxruntime_DELAYLOAD_FLAGS} /DELAYLOAD:${cuda_dll_file_name}")
-    endforeach(cuda_dll_path)
+    # delayload causes crash on exit, so disable for now
+    #file(GLOB cuda_dll_paths "${onnxruntime_CUDA_HOME}/bin/cublas64_*" "${onnxruntime_CUDA_HOME}/bin/cudart64_*" "${onnxruntime_CUDA_HOME}/bin/curand64_*")
+    #set(onnxruntime_DELAYLOAD_FLAGS "${onnxruntime_DELAYLOAD_FLAGS} /DELAYLOAD:cudnn64_7.dll")
+    #foreach(cuda_dll_path ${cuda_dll_paths})
+    #    get_filename_component(cuda_dll_file_name ${cuda_dll_path} NAME)
+    #    set(onnxruntime_DELAYLOAD_FLAGS "${onnxruntime_DELAYLOAD_FLAGS} /DELAYLOAD:${cuda_dll_file_name}")
+    #endforeach(cuda_dll_path)
 
   else()
     link_directories(${onnxruntime_CUDNN_HOME}/lib64)
diff --git a/onnxruntime/test/shared_lib/test_model_loading.cc b/onnxruntime/test/shared_lib/test_model_loading.cc
index ee51301676..b3f88fc2af 100644
--- a/onnxruntime/test/shared_lib/test_model_loading.cc
+++ b/onnxruntime/test/shared_lib/test_model_loading.cc
@@ -2,6 +2,9 @@
 // Licensed under the MIT License.
 
 #include "core/session/onnxruntime_cxx_api.h"
+#ifdef USE_CUDA
+#include "core/providers/cuda/cuda_provider_factory.h"
+#endif
 #include <fstream>
 #include "test_fixture.h"
 #include "file_util.h"
@@ -25,6 +28,12 @@ TEST(CApiTest, model_from_array) {
 
   Ort::SessionOptions so;
   Ort::Session session(*ort_env.get(), buffer.data(), buffer.size(), so);
+
+#ifdef USE_CUDA
+  // test with CUDA provider when using onnxruntime as dll
+  Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_CUDA(so, 0));
+  Ort::Session session_cuda(*ort_env.get(), buffer.data(), buffer.size(), so);
+#endif
 }
 }  // namespace test
 }  // namespace onnxruntime