Simple benchmark is working. Beautiful output.

CMakeLists.txt

@@ -77,6 +77,8 @@ target_compile_definitions(essentials
     ESSENTIALS_VERSION=${ESSENTIALS_VERSION}
 )
 
+message(STATUS "Essentials CUDA Architecture: ${ESSENTIALS_ARCHITECTURES}")
+
 ####################################################
 ############ TARGET COMPILE FEATURES ###############
 ####################################################
@@ -187,6 +189,8 @@ option(ESSENTIALS_BUILD_BENCHMARKS
 
 # Subdirectories for examples, testing and documentation
 if(ESSENTIALS_BUILD_BENCHMARKS)
+  # ... see https://github.com/NVIDIA/nvbench/issues/66
+  set(NVBench_ENABLE_NVML OFF)
   # ... set cuda architecture for nvbench.
   set(CMAKE_CUDA_ARCHITECTURES ${ESSENTIALS_ARCHITECTURES})
   include(${PROJECT_SOURCE_DIR}/cmake/FetchNVBench.cmake)

benchmarks/for.cu

@@ -28,9 +28,12 @@ void parallel_for(nvbench::state& state) {
   cuda::device_id_t device = 0;
   cuda::multi_context_t context(device);
 
-  auto f = [=] __device__(int const& v) -> void { printf("%i\n", v); };
+  vector_t<int> vertices(G.get_number_of_vertices());
+  auto d_vertices = vertices.data().get();
 
-  state.exec([&](nvbench::launch& launch) {
+  auto f = [=] __device__(int const& v) -> void { d_vertices[v] = v; };
+
+  state.exec(nvbench::exec_tag::sync, [&](nvbench::launch& launch) {
     operators::parallel_for::execute<operators::parallel_for_each_t::vertex>(
         G,       // graph
         f,       // lambda function

include/gunrock/container/vector.hxx

@@ -24,7 +24,7 @@ namespace gunrock {
 
 using namespace memory;
 
-template <typename type_t, memory_space_t space>
+template <typename type_t, memory_space_t space = memory_space_t::device>
 using vector_t =
     std::conditional_t<space == memory_space_t::host,  // condition
                        thrust::host_vector<type_t>,    // host_type