Choose a backend#

Every call takes a dyng::resources as its first argument; it decides where the work runs.

auto seq = dyng::resources::sequential();   // one host thread: the reference backend
auto omp = dyng::resources::openmp(16);     // 16 OpenMP threads; 0 = omp_get_max_threads()
auto gpu = dyng::resources::cuda(/*device=*/0);   // GPU 0, its per-thread default stream
auto dflt = dyng::resources();              // dyng::default_backend()

  • sequential is the reference implementation of every algorithm. Use it for tests, for small inputs, and when you need a result to compare against.

  • openmp uses host threads. openmp(0) takes the OpenMP default at the time of the call, which honours OMP_NUM_THREADS; dynG never changes the global OpenMP thread count.

  • cuda runs on one GPU and one stream (resources::cuda(device, stream) takes your own stream_ref). Graphs and results created with it live in device memory: read them with dyng::to_vector(res, tree.distances()). Calls are stream-ordered; each function’s reference entry says when it synchronizes. sssp runs its fused engine, the persistent cooperative kernel of the original MOSP-CUDA code, and on a device without cooperative launch its operators engine (a sequence of kernels with the same results; options::cuda_engine forces either). In a build without CUDA, or with no visible device, resources::cuda() throws not_supported_error.

Ask what this build and machine support before choosing:

if (dyng::backend_available(dyng::backend::openmp)) { /* ... */ }
std::printf("default: %s\n", std::string(dyng::to_string(dyng::default_backend())).c_str());

A backend that was not compiled in throws dyng::not_supported_error when you create its resources, never later in the middle of an algorithm. The deterministic results of an algorithm (for sssp: distances, parents and the deterministic counters) are identical on every backend; each algorithm page states its determinism level.

Copies of a resources object share one handle, so settings made on one copy (attach_profiler, set_memory_resource, set_copy_policy) apply to all of them. Make those settings during setup, and create a separate resources when you need independent settings (Backends and resources).