CUDA: Visual Studio now properly delays device linking

2 files changed, 73 insertions, 7 deletions

diff --git a/Tests/CudaOnly/SeparateCompilation/CMakeLists.txt b/Tests/CudaOnly/SeparateCompilation/CMakeLists.txt
index 420d7a9..0a2542a 100644
--- a/Tests/CudaOnly/SeparateCompilation/CMakeLists.txt
+++ b/Tests/CudaOnly/SeparateCompilation/CMakeLists.txt
@@ -12,6 +12,7 @@ project (CudaOnlySeparateCompilation CUDA)
 string(APPEND CMAKE_CUDA_FLAGS " -gencode arch=compute_30,code=compute_30")
 set(CMAKE_CXX_STANDARD 11)
 set(CMAKE_CUDA_STANDARD 11)
+
 add_library(CUDASeparateLibA STATIC file1.cu file2.cu file3.cu)
 
 #Having file4/file5 in a shared library causes serious problems
@@ -22,12 +23,24 @@ add_library(CUDASeparateLibB STATIC file4.cu file5.cu)
 target_link_libraries(CUDASeparateLibB PRIVATE CUDASeparateLibA)
 
 add_executable(CudaOnlySeparateCompilation main.cu)
-target_link_libraries(CudaOnlySeparateCompilation PRIVATE CUDASeparateLibB)
+target_link_libraries(CudaOnlySeparateCompilation
+                      PRIVATE CUDASeparateLibB)
+
+set_target_properties(CUDASeparateLibA
+                      CUDASeparateLibB
+                      PROPERTIES CUDA_SEPARABLE_COMPILATION ON
+                      POSITION_INDEPENDENT_CODE ON)
 
-set_target_properties( CUDASeparateLibA
-                       CUDASeparateLibB
-                       PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
+if (CMAKE_GENERATOR MATCHES "^Visual Studio")
+  #Visual Studio CUDA integration will not perform device linking
+  #on a target that itself does not have GenerateRelocatableDeviceCode
+  #enabled.
+  set_target_properties(CudaOnlySeparateCompilation
+                        PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
+endif()
 
-set_target_properties( CUDASeparateLibA
-                       CUDASeparateLibB
-                       PROPERTIES POSITION_INDEPENDENT_CODE ON)
+if (APPLE)
+  # We need to add the default path to the driver (libcuda.dylib) as an rpath, so that
+  # the static cuda runtime can find it at runtime.
+  target_link_libraries(CudaOnlySeparateCompilation PRIVATE -Wl,-rpath,/usr/local/cuda/lib)
+endif()
diff --git a/Tests/CudaOnly/SeparateCompilation/main.cu b/Tests/CudaOnly/SeparateCompilation/main.cu
index 03e0921..40dbe5d 100644
--- a/Tests/CudaOnly/SeparateCompilation/main.cu
+++ b/Tests/CudaOnly/SeparateCompilation/main.cu
@@ -7,9 +7,62 @@
 int file4_launch_kernel(int x);
 int file5_launch_kernel(int x);
 
+int choose_cuda_device()
+{
+  int nDevices = 0;
+  cudaError_t err = cudaGetDeviceCount(&nDevices);
+  if (err != cudaSuccess) {
+    std::cerr << "Failed to retrieve the number of CUDA enabled devices"
+              << std::endl;
+    return 1;
+  }
+  for (int i = 0; i < nDevices; ++i) {
+    cudaDeviceProp prop;
+    cudaError_t err = cudaGetDeviceProperties(&prop, i);
+    if (err != cudaSuccess) {
+      std::cerr << "Could not retrieve properties from CUDA device " << i
+                << std::endl;
+      return 1;
+    }
+    if (prop.major >= 3) {
+      err = cudaSetDevice(i);
+      if (err != cudaSuccess) {
+        std::cout << "Could not select CUDA device " << i << std::endl;
+      } else {
+        return 0;
+      }
+    }
+  }
+
+  std::cout << "Could not find a CUDA enabled card supporting compute >=3.0"
+            << std::endl;
+
+  return 1;
+}
+
 int main(int argc, char** argv)
 {
+  int ret = choose_cuda_device();
+  if (ret) {
+    return 0;
+  }
+
+  cudaError_t err;
   file4_launch_kernel(42);
+  err = cudaGetLastError();
+  if (err != cudaSuccess) {
+    std::cerr << "file4_launch_kernel: kernel launch failed: "
+              << cudaGetErrorString(err) << std::endl;
+    return 1;
+  }
+
   file5_launch_kernel(42);
+  err = cudaGetLastError();
+  if (err != cudaSuccess) {
+    std::cerr << "file5_launch_kernel: kernel launch failed: "
+              << cudaGetErrorString(err) << std::endl;
+    return 1;
+  }
+
   return 0;
 }