Heterogeneous many-core-oriented data transmission method based on space rearrangement

Abstract

The invention discloses a heterogeneous many-core-oriented space rearrangement-based data transmission method, which comprises the following steps of S1, during compiling, establishing a data transmission information table for each acceleration area, the data transmission information table comprising information of to-be-transmitted data of a current acceleration area; S2, according to a data transmission mode, performing spatial rearrangement on multiple pieces of to-be-transmitted data; S3, before calculation of the acceleration area is started, starting DMA data transmission from main storage data to a local storage buffer area, and transmitting data needed by calculation of the acceleration area, including read-in data and read-write data, from a main storage to the local storage buffer area at a time; and S4, after the calculation of the acceleration area is completed, starting DMA data transmission from the local memory buffer area to a main memory and transmitting data needing to be transmitted by the acceleration area and needing to be written out, including read-write data and written-out data, to the main memory from the local memory buffer area at a time. Expenditure ofmultiple data transmission between the many-core on-chip local memory and the main memory is effectively reduced, and the memory access efficiency and the overall performance of a program are improved.

Description

technical field [0001] The invention relates to a data transmission method based on spatial rearrangement for heterogeneous many cores, and belongs to the technical field of computers. Background technique [0002] Under the heterogeneous many-core multi-level storage architecture, data transmission between on-chip local memory and main memory is essential, and the efficiency of data transmission has a great impact on the performance of many-core applications. Data is generally transferred between the local memory on the many-core chip and the main memory through DMA (Direct Memory Access). If a block of data is used in the acceleration area of ​​many-core applications, a DMA operation is initiated before accessing the block of data, and the block of data is copied from the main memory to the on-chip local memory. If multiple blocks of data are used, a DMA operation is initiated for each block of data. [0003] However, the existing on-chip data transfer method between the...

AI Technical Summary

Problems solved by technology

[0003] However, the existing on-chip data transfer method between the local memory and the main memory has the following defects: first, if there are multiple blocks of data to be transferred, multiple DMA operations will be initiated, causing excessive data transfer startup overhead; However, if multiple small data blocks are to be transferred, the amount of data transferred by each DMA operation is very small, which cannot make full use of the data transfer bandwidth, reduces data transfer efficiency, and affects program performance

Images