Can this operation change the value of grad_input

Tensor grad_input_n = grad_input.select(0, elt);

        col2hvol<scalar_t, dim>(
            stream,
            columns.data_ptr<scalar_t>(),
            nInputPlane,
            input_size,
            output_size,
            kernel_size,
            stride_size,
            pad_size,
            dilation_size,
            grad_input_n.data_ptr<scalar_t>());
      }

file “pytorch\aten\src\ATen\native\cuda\NaiveDilatedConvolution.cu”

Hi,

Yes, from what I understand, it copies the results from the gemm stored in columns back into the grad_input Tensor that will be returned as the gradient wrt the input.