parallel memory modules can be used to increase memory bandwidth and feed a processor with the required access patterns of data. The parallelstorage mechanism organized and managed by multiple storage modules can sui...
详细信息
ISBN:
(纸本)9783030050573;9783030050566
parallel memory modules can be used to increase memory bandwidth and feed a processor with the required access patterns of data. The parallelstorage mechanism organized and managed by multiple storage modules can suit applications of images and videos. Previous investigation into data storageschemes can be used to achieve continuous conflict free access by rows, columns or blocks, however it is not only satisfied with some sliding window applications in video and image processing algorithms (including convolutional neural networks, sub-pixel difference, 2D filtering, etc.) which need non-conflicting access by steps in computation, but also there is a different demand for horizontal and vertical strides in computing sub-processes. This paper presents a storagescheme that support for row access without collision alignment, and non-aligned block-with-stride access storage modes beginning at any address. Theoretical proofs and experiments verify the correct ness of the module address (module number to which the address is mapped). And in hardware design, it was found that in the typical case there was no path violation and with less area overhead. It suitable for application of CNN to improve performance in algorithm in convolutional.
暂无评论