“…In addition, the SFG of LLM for forward and inverse DCTs are mirror images of one another and hence we can re-use the entire datapath for performing the forward and inverse transforms. As can be seen from Fig.1, for subsampled blocks of sizes 4x (1,2,4,8), we perform only the upper half computation(as indicated by the dashed line box) and for subsampled blocks of sizes 2x (1,2,4,8) we perform only the upper quarter computation(as indicated by the solid-line box). This is because the other coefficients are zero.…”