Parallel discrete updating method for heterogeneous many-core processor

Abstract

The invention discloses a parallel discrete updating method for a heterogeneous many-core processor, which comprises the following steps of: S1, carrying out block division on a discrete updating data area with continuous addresses, and mapping data blocks and slave cores, and therefore each slave core is responsible for updating one data block; S2, enabling each slave core to classify the discrete updating requests generated in the instruction execution process into discrete updating requests needing to be processed by the slave core and discrete updating requests needing to be processed by other slave cores, and inserting corresponding processing functions into the positions where the discrete updating requests are generated; S3, processing a discrete updating request which needs to be processed by the slave core according to the processing flow of the program segment func1; S4, processing the discrete updating requests which need to be processed by other slave cores according to the processing flow of the program segment func2. According to the method, high-performance parallel discrete updating can be realized, and the problem of discrete updating of big data which cannot be installed in an LDM space can be solved.

Description

technical field [0001] The invention belongs to the technical field of computer compilation and optimization, in particular to a parallel discrete update method for heterogeneous many-core processors. Background technique [0002] There are two types of discrete updates: one is write-before-read, indicating that there is a correlation between loop iterations, and such loops usually cannot be parallelized; the other is write-related, as long as the atomicity of the write operation is guaranteed, that is, Can be parallelized. [0003] Usually, the producer-consumer model is adopted on the heterogeneous many-core platform to solve the discrete update problem: the slave core is divided into two parts: the computing core (producer) and the update core (consumer), and the producer is responsible for the discrete update request (the middle Calculation results), determine which consumer the update request should belong to, and send it to the corresponding consumer through the commu...

AI Technical Summary

Problems solved by technology

[0004] The disadvantages of the producer-consumer mode technology include: 1) The source code must be transformed according to the requirements of the producer-consumer mode, and a simple segment of the array discrete update cycle code segment in the original code is transformed into a complex producer-consumer code segment , it is not convenient for the compiler to automatically transform the program, which reduces the readability and portability of the program; 2) It is necessary to divide the slave cores in the slave core group into two parts in proportion: producers and consumers. This ratio requires It is determined by weighing factors such as core processing power and computing load, or obtained by running tests, which increases the programming burden

Images