Question: Why does scaled_back_data shared across samples within a batch?

[hzxie]

bilateralNN/bilateralnn_code/src/caffe/layers/permutohedral_layer.cu

Lines 134 to 148 in e174e4e

	Dtype* scaled_back_data = scaled_back.mutable_gpu_data();
	const int in_size = in_height_ * in_width_;
	const int out_size = out_height_ * out_width_;
	for (int n = 0; n < num_; ++n) {
	BlurOperation& op = operations_[n];
	const Dtype* norm_there_data = op.norm_there_->gpu_data();
	const Dtype* norm_back_data = op.norm_back_->gpu_data();
	for (int c = 0; c < num_output_; ++c) {
	caffe_gpu_mul(out_size,
	top_diff + top_blob.offset(n, c), norm_back_data,
	scaled_back_data + scaled_back.offset(0, c));

According to the code above, the scaled_back_data is shared across samples within a batch?
Could you tell me the reason?

I think the scaled_back_data should be decalred within the for loop.

[hzxie]

Because the variable is used to save the result for the current sample.