CUDA: Add test for device linking when host linking uses threads

Convert the `CudaOnly.LinkSystemDeviceLibraries` test to a new
`Cuda.ProperDeviceLibraries` test.  The former covered only the
`cublas_device` library which is removed by CUDA 10.  Extend the new
test to also cover various cases of using threads.

Issue: #18008

3 files changed, 0 insertions, 99 deletions

diff --git a/Tests/CudaOnly/CMakeLists.txt b/Tests/CudaOnly/CMakeLists.txt
index 5b7c0e6..9c4f86a 100644
--- a/Tests/CudaOnly/CMakeLists.txt
+++ b/Tests/CudaOnly/CMakeLists.txt
@@ -3,7 +3,6 @@ ADD_TEST_MACRO(CudaOnly.CircularLinkLine CudaOnlyCircularLinkLine)
 ADD_TEST_MACRO(CudaOnly.EnableStandard CudaOnlyEnableStandard)
 ADD_TEST_MACRO(CudaOnly.ExportPTX CudaOnlyExportPTX)
 ADD_TEST_MACRO(CudaOnly.GPUDebugFlag CudaOnlyGPUDebugFlag)
-ADD_TEST_MACRO(CudaOnly.LinkSystemDeviceLibraries CudaOnlyLinkSystemDeviceLibraries)
 ADD_TEST_MACRO(CudaOnly.ResolveDeviceSymbols CudaOnlyResolveDeviceSymbols)
 ADD_TEST_MACRO(CudaOnly.SeparateCompilation CudaOnlySeparateCompilation)
 ADD_TEST_MACRO(CudaOnly.WithDefs CudaOnlyWithDefs)
diff --git a/Tests/CudaOnly/LinkSystemDeviceLibraries/CMakeLists.txt b/Tests/CudaOnly/LinkSystemDeviceLibraries/CMakeLists.txt
deleted file mode 100644
index 7f7f606..0000000
--- a/Tests/CudaOnly/LinkSystemDeviceLibraries/CMakeLists.txt
+++ /dev/null
@@ -1,15 +0,0 @@
-cmake_minimum_required(VERSION 3.8)
-project(LinkSystemDeviceLibraries CUDA)
-
-string(APPEND CMAKE_CUDA_FLAGS " -gencode arch=compute_35,code=compute_35 -gencode arch=compute_35,code=sm_35")
-set(CMAKE_CUDA_STANDARD 11)
-
-add_executable(CudaOnlyLinkSystemDeviceLibraries main.cu)
-set_target_properties( CudaOnlyLinkSystemDeviceLibraries
-                       PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
-target_link_libraries( CudaOnlyLinkSystemDeviceLibraries PRIVATE cublas_device)
-
-if(APPLE)
-  # Help the static cuda runtime find the driver (libcuda.dyllib) at runtime.
-  set_property(TARGET CudaOnlyLinkSystemDeviceLibraries PROPERTY BUILD_RPATH ${CMAKE_CUDA_IMPLICIT_LINK_DIRECTORIES})
-endif()
diff --git a/Tests/CudaOnly/LinkSystemDeviceLibraries/main.cu b/Tests/CudaOnly/LinkSystemDeviceLibraries/main.cu
deleted file mode 100644
index 2c7c388..0000000
--- a/Tests/CudaOnly/LinkSystemDeviceLibraries/main.cu
+++ /dev/null
@@ -1,83 +0,0 @@
-
-#include <cublas_v2.h>
-#include <cuda_runtime.h>
-#include <iostream>
-
-// this test only makes sense for versions of CUDA that ships
-// static libraries that have separable compilation device symbols
-#if __CUDACC_VER_MAJOR__ <= 9
-__global__ void deviceCublasSgemm(int n, float alpha, float beta,
-                                  const float* d_A, const float* d_B,
-                                  float* d_C)
-{
-  cublasHandle_t cnpHandle;
-  cublasStatus_t status = cublasCreate(&cnpHandle);
-
-  if (status != CUBLAS_STATUS_SUCCESS) {
-    return;
-  }
-
-  // Call function defined in the cublas_device system static library.
-  // This way we can verify that we properly pass system libraries to the
-  // device link line
-  status = cublasSgemm(cnpHandle, CUBLAS_OP_N, CUBLAS_OP_N, n, n, n, &alpha,
-                       d_A, n, d_B, n, &beta, d_C, n);
-
-  cublasDestroy(cnpHandle);
-}
-#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 || (prop.major == 3 && prop.minor >= 5)) {
-      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 >=3.5"
-            << std::endl;
-  return 1;
-}
-
-int main(int argc, char** argv)
-{
-  int ret = choose_cuda_device();
-  if (ret) {
-    return 0;
-  }
-
-#if __CUDACC_VER_MAJOR__ <= 9
-  // initial values that will make sure that the cublasSgemm won't actually
-  // do any work
-  int n = 0;
-  float alpha = 1;
-  float beta = 1;
-  float* d_A = nullptr;
-  float* d_B = nullptr;
-  float* d_C = nullptr;
-  deviceCublasSgemm<<<1, 1>>>(n, alpha, beta, d_A, d_B, d_C);
-#endif
-
-  return 0;
-}