The routine does no error checking and assumes that both A and B matrix and for the calling matrix, the sizes, FloatPrecision, the layout and the memory location (!) do not change between calls. The matrices multiplied are those passed to the MulSmallInit routine. You can change the values in the memory allocated by those matrices, but may not resize them or resize the calling matrix. 

The kernel for the particular configuration is initialized with a call to MulSmallInit and released either when a new one is created or when the calling matrix is freed.

MulSmall Method