Function ForEachFast

oneflow/user/kernels/normalization_kernel.cpp:228–237

// NOTE(Liang Depeng): helper functions to process datas for specific channel over all samples.
template<typename T, typename DataProcessor>
static inline void ForEachFast(const T* data, const int64_t batch_size, const int64_t spatial_size,
                               const int64_t jump_step, const int64_t channel_idx,
                               DataProcessor data_processor) {
  const int64_t start_offset = channel_idx * spatial_size;
  const T* tmp_data = data + start_offset;
  for (int64_t outer = 0; outer < batch_size; ++outer) {
    for (int64_t i = 0; i < spatial_size; ++i) { data_processor(&tmp_data[i]); }
    tmp_data += jump_step;
  }
}

template<typename T, typename DataProcessor>
static inline void ForEachFast(const T* in_data1, const T* in_data2, const int64_t batch_size,