Many recent block ciphers use Maximum Distance Separable (MDS) matrices in their diffusion layer. The main objective of this operation is to spread as much as possible the differences between the outputs of nonlinear Sboxes. So they generally act at nibble or at byte level. The MDS matrices are associated to MDS codes of ratio 1/2. The most famous example is the MixColumns operation of the AES block cipher. In this example, the MDS matrix was carefully chosen to obtain compact and efficient implementations. However, this MDS matrix is dedicated to 8-bit words, and is not always adapted to lightweight applications. Recently, several studies have been devoted to the construction of recursive diffusion layers. Such a method allows to apply an MDS matrix using an iterative process which looks like a Feistel network with linear functions instead of nonlinear. Our approach is quite different. We present a generic construction of classical MDS matrices that are not recursively computed, but that are strong symmetric in order to either accelerate their evaluation with a minimal number of look-up tables, or to perform this evaluation with a minimal number of gates in a circuit. We call this particular kind of matrices "dyadic matrices", since they are related to dyadic codes. We study some basic properties of such matrices. We introduce a generic construction of involutive dyadic MDS matrices from Reed Solomon codes. Finally, we discuss the implementation aspects of these dyadic MDS matrices in order to build efficient block ciphers.