Method ParallelForImpl

oneflow/core/thread/thread_runtime.h:99–109

class TbbRuntime final : public RuntimeBase {
 private:
  void ParallelForImpl(int64_t begin, int64_t end, const CallableT& func, size_t num_threads,
                       size_t grain_size) override {
    tbb::global_control global_thread_limit(tbb::global_control::max_allowed_parallelism,
                                            num_threads);
    const size_t chunk_size = std::max(DivUp((end - begin), num_threads), grain_size);

    tbb::parallel_for(
        tbb::blocked_range<int64_t>(begin, end, chunk_size),
        [&func](const tbb::blocked_range<int64_t>& r) { SeqFor(r.begin(), r.end(), func); },
        tbb::static_partitioner{});
  }
};
#endif