Thread affinity controls where OpenMP threads run. FDTD CUDA's team helper selects explicit binding when the variant requests it:

if (bind == OmpBind::Close) {
    #pragma omp parallel num_threads(threads) proc_bind(close)
    {
        body();
    }
} else if (bind == OmpBind::Spread) {
    #pragma omp parallel num_threads(threads) proc_bind(spread)
    {
        body();
    }
}

The +close and +spread variant suffixes select these branches. For example:

$ OMP_PLACES=cores build/out/debug/fdtd \
    --nx 512 --ny 512 --steps 400 \
    --backend openmp --variant v4+spread

OMP_PLACES defines the available processor locations. Close binding favors nearby places; spread partitions them more widely. Without a suffix, the backend leaves binding to runtime configuration. Compare policies under the same workload: placement can affect shared caches and memory access, but neither policy guarantees a speedup.

Source: fdtd_tmz_backend_omp_team.h.