A system and method for reading and writing compatible two-dimensional structured data

Abstract

A compatible two-dimensional structure data reading and writing system, comprising: a random access memory selection unit, which reads and writes SRAM according to the reading and writing requirements of the instruction decoding of the operation execution unit; a predictive cache unit, which predicts the two-dimensional data required for instruction execution Cache; write strategy unit, it is used to formulate the strategy that data is written into SRAM and writes back external storage; Cache tag cache unit, it caches the valid mark of Cache, the state value of square; Cache preload unit, it according to described Cache marks the update flag of the cache unit, generates a corresponding block refresh request, and puts it into the DMA unit request queue to update the Cache; the pre-read unit analyzes the computing tasks according to the preset batch, and marks the tasks . The system of the invention greatly improves the computing efficiency of the system with vector processing capability.

Description

technical field [0001] The invention relates to the technical field of data processing, in particular to a system and method for reading and writing compatible two-dimensional structure data. Background technique [0002] In the digital logic system, in order to balance the relationship between the access of the slower storage unit and the high-speed operation processing unit, it is often necessary to introduce a Cache (caching) design. For data sources that exist in peripheral memory (DDR external memory, etc.) in two-dimensional form, such as digital images, sensor acquisition data, and neural network data stored in a similar form that can be equivalent to a large matrix, the algorithm logic often needs to be obtained one by one. A square area with m rows and n columns (or vector form, which can be regarded as the distortion of a matrix) of a two-dimensional data file or a two-dimensional graphic area of ​​other shapes is used for scientific operations such as convolution ...

AI Technical Summary

Problems solved by technology

For data sources that exist in peripheral memory (DDR external memory, etc.) in two-dimensional form, such as digital images, sensor acquisition data, and neural network data stored in a similar form that can be equivalent to a large matrix, the algorithm logic often needs to be obtained one by one. A square area with m rows and n columns (or vector form, which can be regarded as the distortion of a matrix) of a two-dimensional data file or a two-dimensional graphic area of ​​other shapes is used for scientific operations such as convolution operations and iterative operations, while This kind of unit area with multiple consecutive calculations has a certain spatial and time domain continuity that is displaced from the center point of the area to the surroundings, so the unit area will have considerable overlap between consecutive multiple operations (such as figure 1 ), intuitively, this requires that when reading data from external memory to refresh the cache, the overlapping part should be kept in the cache as long as possible, and the unit area can be acquired by the operation processing unit faster, otherwise it will slow down the system Actual operation and processing efficiency

Images