The implementation in this paper packs 4 cons 4*1 consecutive elements from a matrix column into a texel and thus performs a small number of 4*4 matrix by 4*1 vector products in a shader.

• 每次从矩阵中取出4个元素放入一个到文理单元的一个元素中，4×4的小规模矩阵乘法就转变成了4×1的向量乘法。