Method Initialize

Src/Base/AMReX_GpuDevice.cpp:225–1408

#endif

void
Device::Initialize (bool minimal, int a_device_id)
{
    amrex::ignore_unused(minimal, a_device_id);
#ifdef AMREX_USE_GPU

#if defined(AMREX_USE_CUDA) && (defined(AMREX_PROFILING) || defined(AMREX_TINY_PROFILING))
    // Wrap cuda init to identify it appropriately in nvvp.
    // Note: first substantial cuda call may cause a lengthy
    // cuda API and cuda driver API initialization that will
    // be captured by the profiler. It a necessary, system
    // dependent step that is unavoidable.
    nvtxRangePush("initialize_device");
#endif

    ParmParse ppamrex("amrex");
    ppamrex.queryAdd("max_gpu_streams", max_gpu_streams);
    max_gpu_streams = std::min(max_gpu_streams, AMREX_GPU_MAX_STREAMS);
    max_gpu_streams = std::max(max_gpu_streams, 1);

    ParmParse pp("device");
    if (! pp.query("verbose", "v", verbose)) {
        pp.add("verbose", verbose);
    }

    if (amrex::Verbose()) {
        AMREX_HIP_OR_CUDA_OR_SYCL
            ( amrex::Print() << "Initializing HIP...\n";,
              amrex::Print() << "Initializing CUDA...\n";,
              amrex::Print() << "Initializing SYCL...\n"; )
    }

    // Count the number of GPU devices.
    int gpu_device_count = 0;
#ifdef AMREX_USE_SYCL
    {
        sycl::platform platform(sycl::gpu_selector_v);
        auto const& gpu_devices = platform.get_devices();
        gpu_device_count = gpu_devices.size();
        if (gpu_device_count <= 0) {
            amrex::Abort("No GPU device found");
        }
    }
#else
    AMREX_HIP_OR_CUDA(AMREX_HIP_SAFE_CALL (hipGetDeviceCount(&gpu_device_count));,
                      AMREX_CUDA_SAFE_CALL(cudaGetDeviceCount(&gpu_device_count)); );
    if (gpu_device_count <= 0) {
        amrex::Abort("No GPU device found");
    }
#endif

    // Now, assign ranks to GPUs. If we only have one GPU,
    // or only one MPI rank, this is easy. Otherwise, we
    // need to do a little more work.

    int n_local_procs = 1;
    amrex::ignore_unused(n_local_procs);