Method InspectNumaNodes

tensorflow/stream_executor/cuda/cuda_platform.cc:78–100

Due to legacy issues in user code, we can't currently call InpectNumaNodes at module initialization time, because non-GPU programs still include this plugin via various methods, so instead, it has to be init-on-reference.

// at module initialization time, because non-GPU programs still include this
// plugin via various methods, so instead, it has to be init-on-reference.
void CudaPlatform::InspectNumaNodes() {
  // To get NUMA node information, we need to create all executors, so we can
  // examine their device descriptions to see their bus assignments.
  static std::once_flag once;
  std::call_once(once, [&] {
    StreamExecutorConfig config;
    for (int i = 0; i < VisibleDeviceCount(); i++) {
      config.ordinal = i;
      StreamExecutor* exec = GetExecutor(config).ValueOrDie();
      if (i == 0) {
        // NUMA nodes may not start at 0, so set the minimum node  based on the
        // first executor we see.
        min_numa_node_ = exec->GetDeviceDescription().numa_node();
        limit_numa_node_ = min_numa_node_ + 1;
      } else {
        min_numa_node_ =
            std::min(min_numa_node_, exec->GetDeviceDescription().numa_node());
        limit_numa_node_ = std::max(
            limit_numa_node_, exec->GetDeviceDescription().numa_node() + 1);
      }
    }
  });
}

int CudaPlatform::BusCount() {
  InspectNumaNodes();