DEVICE AND METHOD FOR ACCELERATING CONSISTENT CONVERSION BETWEEN INTER-DATA EXPRESSION

摘要

To provide a device and a method for converting a set of source tensor block having a plurality of source tensor data elements including at least one first numerical value expression to at least one destination tensor block in the sequence specified on the basis of the first numerical value expression and a second numerical value expression.SOLUTION: The method includes the steps of: fetching a source tensor block of a source tensor data structure; converting at least one source tensor block to at least one destination tensor block having a plurality of destination tensor data elements including a second numerical value expression different from a first numerical value expression; and storing each destination tensor block into the specified memory region to maintain the consistency with the arrangement structure defined in advance of the source tensor block.SELECTED DRAWING: Figure 25