Method BeforeFindBestSplit

src/treelearner/gpu_tree_learner.cpp:821–870

}

bool GPUTreeLearner::BeforeFindBestSplit(const Tree* tree, int left_leaf, int right_leaf) {
  int smaller_leaf;
  data_size_t num_data_in_left_child = GetGlobalDataCountInLeaf(left_leaf);
  data_size_t num_data_in_right_child = GetGlobalDataCountInLeaf(right_leaf);
  // only have root
  if (right_leaf < 0) {
    smaller_leaf = -1;
  } else if (num_data_in_left_child < num_data_in_right_child) {
    smaller_leaf = left_leaf;
  } else {
    smaller_leaf = right_leaf;
  }

  // Copy indices, gradients and hessians as early as possible
  if (smaller_leaf >= 0 && num_dense_feature_groups_) {
    // only need to initialize for smaller leaf
    // Get leaf boundary
    const data_size_t* indices = data_partition_->indices();
    data_size_t begin = data_partition_->leaf_begin(smaller_leaf);
    data_size_t end = begin + data_partition_->leaf_count(smaller_leaf);

    // copy indices to the GPU:
    #if GPU_DEBUG >= 2
    Log::Info("Copying indices, gradients and hessians to GPU...");
    printf("Indices size %d being copied (left = %d, right = %d)\n", end - begin, num_data_in_left_child, num_data_in_right_child);
    #endif
    indices_future_ = boost::compute::copy_async(indices + begin, indices + end, device_data_indices_->begin(), queue_);

    if (!is_constant_hessian_) {
      #pragma omp parallel for schedule(static)
      for (data_size_t i = begin; i < end; ++i) {
        ordered_hessians_[i - begin] = hessians_[indices[i]];
      }
      // copy ordered hessians to the GPU:
      hessians_future_ = queue_.enqueue_write_buffer_async(device_hessians_, 0, (end - begin) * sizeof(score_t), ptr_pinned_hessians_);
    }

    #pragma omp parallel for schedule(static)
    for (data_size_t i = begin; i < end; ++i) {
      ordered_gradients_[i - begin] = gradients_[indices[i]];
    }
    // copy ordered gradients to the GPU:
    gradients_future_ = queue_.enqueue_write_buffer_async(device_gradients_, 0, (end - begin) * sizeof(score_t), ptr_pinned_gradients_);

    #if GPU_DEBUG >= 2
    Log::Info("Gradients/hessians/indices copied to device with size %d", end - begin);
    #endif
  }
  return SerialTreeLearner::BeforeFindBestSplit(tree, left_leaf, right_leaf);
}

bool GPUTreeLearner::ConstructGPUHistogramsAsync(
  const std::vector<int8_t>& is_feature_used,