CNN convolution fast implementation

Below is my implementation of the CNN convolution. "in" is a 3D matrix, there are also a 4D matrix "weights" and a 1D matrix "bias". "Out" is a 3D matrix. I am wondering if there is a better fast implementation.

function out = spConvLayerForward(in, layer, weights, bias)
[row, col, planeIn] = size(in);
planeOut = size(layer.weights, 4);

out = single(zeros(row, col, planeOut));

for i = 1 : planeOut
    bias = layer.bias(i);
    tmp = out(:, :, i);
    tmp(:) = 0;
    for j = 1 : planeIn
        tmp = tmp + conv2(in(:, :, j), weight, 'same');
    end
    tmp = tmp + bias;
end

[Walter Roberson]

You can reshape in() to 2d and use a single conv2() call, but the boundary conditions become a nuisance.

Where is the variable "weight" coming from, and why did you make it so similar to the unused parameter "weights"?