An apparatus for performing a method for loading a matrix into an accelerator includes an augmented direct memory access controller reading a matrix, in a data stream, from a first memory associated with a system processor and sending the matrix, in the data stream, to a second memory associated with the accelerator. The method further includes the augmented direct memory access controller extracting individual matrix elements from the data stream as the data stream is being sent to the second memory and analyzing the extracted individual matrix elements to determine if the matrix is any of a plurality of tested matrix class types as the data stream is being sent to the second memory.