CPC G06F 13/1668 (2013.01) [G06N 3/04 (2013.01)] | 19 Claims |
1. A method of operating an accelerator comprising an input memory and configured to perform a target operation, the input memory comprising words for storing, at one time, input data comprising a data map, the words having a same word width, the words including a word, the method comprising:
determining a data layout in accordance with striding a filter over the data map such that a first portion of the data map corresponding to a first stride of the filter partially overlaps a second portion of the data map corresponding to a second stride of the filter that partially overlaps the first stride of the filter, and packing the first portion of the data map and a part of the second portion that does not overlap the first portion into the word of the input memory at the same time according to the data layout, wherein the data layout is determined based on the word width and a spatial size of the filter, which is to be applied to the target operation, wherein the data layout is based on a number of partially overlapping portions of the input data to be packed into one word; and
performing the target operation between the first portion of the data map packed in the word in the memory and weights of the filter.
|