CUDA: Allow -G to control device debuging on MSVC.

Fixes #17551

3 files changed, 91 insertions, 1 deletions

diff --git a/Tests/CudaOnly/CMakeLists.txt b/Tests/CudaOnly/CMakeLists.txt
index 5f456fc..5ad6e6b 100644
--- a/Tests/CudaOnly/CMakeLists.txt
+++ b/Tests/CudaOnly/CMakeLists.txt
@@ -1,6 +1,7 @@
 
 ADD_TEST_MACRO(CudaOnly.EnableStandard CudaOnlyEnableStandard)
 ADD_TEST_MACRO(CudaOnly.ExportPTX CudaOnlyExportPTX)
+ADD_TEST_MACRO(CudaOnly.GPUDebugFlag CudaOnlyGPUDebugFlag)
+ADD_TEST_MACRO(CudaOnly.ResolveDeviceSymbols CudaOnlyResolveDeviceSymbols)
 ADD_TEST_MACRO(CudaOnly.SeparateCompilation CudaOnlySeparateCompilation)
 ADD_TEST_MACRO(CudaOnly.WithDefs CudaOnlyWithDefs)
-ADD_TEST_MACRO(CudaOnly.ResolveDeviceSymbols CudaOnlyResolveDeviceSymbols)
diff --git a/Tests/CudaOnly/GPUDebugFlag/CMakeLists.txt b/Tests/CudaOnly/GPUDebugFlag/CMakeLists.txt
new file mode 100644
index 0000000..5b96906
--- /dev/null
+++ b/Tests/CudaOnly/GPUDebugFlag/CMakeLists.txt
@@ -0,0 +1,23 @@
+
+cmake_minimum_required(VERSION 3.7)
+project (CudaOnlGPUDebugFlag CUDA)
+
+#Goal for this example:
+#verify that -G enables gpu debug flags
+string(APPEND CMAKE_CUDA_FLAGS " -gencode=arch=compute_30,code=compute_30")
+string(APPEND CMAKE_CUDA_FLAGS " -G")
+set(CMAKE_CUDA_STANDARD 11)
+
+add_executable(CudaOnlyGPUDebugFlag main.cu)
+
+if(CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 9.0.0)
+  #CUDA's __CUDACC_DEBUG__ define was added in 9.0
+  #so if we are below 9.0.0 we will manually add the define so that the test
+  #passes
+  target_compile_definitions(CudaOnlyGPUDebugFlag PRIVATE "__CUDACC_DEBUG__")
+endif()
+
+if(APPLE)
+  # Help the static cuda runtime find the driver (libcuda.dyllib) at runtime.
+  set_property(TARGET CudaOnlyGPUDebugFlag PROPERTY BUILD_RPATH ${CMAKE_CUDA_IMPLICIT_LINK_DIRECTORIES})
+endif()
diff --git a/Tests/CudaOnly/GPUDebugFlag/main.cu b/Tests/CudaOnly/GPUDebugFlag/main.cu
new file mode 100644
index 0000000..8b97a3f
--- /dev/null
+++ b/Tests/CudaOnly/GPUDebugFlag/main.cu
@@ -0,0 +1,66 @@
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include <iostream>
+
+static __global__ void debug_kernel(bool* has_debug)
+{
+// Verify using the return code if we have GPU debug flag enabled
+#if defined(__CUDACC__) && defined(__CUDACC_DEBUG__)
+  *has_debug = true;
+#else
+  *has_debug = false;
+#endif
+}
+
+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)
+{
+  bool* has_debug;
+  cudaError_t err = cudaMallocManaged(&has_debug, sizeof(bool));
+
+  debug_kernel<<<1, 1>>>(has_debug);
+  err = cudaDeviceSynchronize();
+  if (err != cudaSuccess) {
+    std::cerr << "debug_kernel: kernel launch shouldn't have failed\n"
+              << "reason:\t" << cudaGetErrorString(err) << std::endl;
+    return 1;
+  }
+  if (*has_debug == false) {
+    std::cerr << "debug_kernel: kernel not compiled with device debug"
+              << std::endl;
+    return 1;
+  }
+  return 0;
+}