int64_t total_us;                  // Total time in us.

  Stats() : total_us(0) { per_iter_us.reserve(5000); }
};

// DumpStatsToStdout printfs to stdout stats in a multi-line human-friendly
// form.
void DumpStatsToStdout(const Stats& stats);

// BenchmarkFn is the signature of the function generated by tfcompile.
typedef std::function<void()> BenchmarkFn;

  CPP_CLASS computation;
  computation.set_thread_pool(&device);

  benchmark::Options options;
  benchmark::Stats stats;
  benchmark::Benchmark(options, [&] { computation.Run(); }, &stats);
  benchmark::DumpStatsToStdout(stats);
  return 0;
}

}  // namespace tfcompile
}  // namespace tensorflow

int main(int argc, char** argv) {
  return tensorflow::tfcompile::Main(argc, argv);

static uint64 NowMicros() {
  struct timeval tv;
  gettimeofday(&tv, nullptr);
  return static_cast<uint64>(tv.tv_sec) * 1000000 + tv.tv_usec;
}

void DumpStatsToStdout(const Stats& stats) {
  // Compute stats.
  std::vector<int64_t> sorted_us(stats.per_iter_us);
  std::sort(sorted_us.begin(), sorted_us.end());
  const size_t count_us = sorted_us.size();
  double sum_us = 0;