Implicit data dynamic reuse method for many-core distributed local storage

Abstract

The invention discloses an implicit data dynamic reusing method oriented to many-core distributed local storage, which comprises the following steps of: S1, respectively processing a storage mode of data in a local local storage of an acceleration calculation core according to a data access mode and a data volume in the acceleration calculation and the capacity of the local local storage of the acceleration calculation core; s2, marking data variable names or array offsets which may be reused in a program through registration compiling indication; s3, creating a main storage address and local storage address mapping table of the reused data; s4, indicating a data variable name or array offset which needs to be reused through a reusing compilation indication; and S6, performing replacement processing on the access to the reused data main storage variable in the code segment marked by the'reuse compiling indication '. According to the method, release can be dynamically applied, the limited local storage space is fully utilized, the reused data can reside in the local storage as long as possible, the overhead of data transmission is reduced, and the performance of a program is improved.

Description

technical field [0001] The invention relates to an implicit data dynamic reuse method for many-core distributed local storage, and belongs to the technical field of high-performance computing. Background technique [0002] Many-core systems generally have multi-level storage systems, including off-chip main memory and each or every few computing cores have a distributed local memory with independent addressing, and the local memory is completely managed by the user or software. The capacity of the main memory is large, while the capacity of the local memory is small. Compared with the main memory, the accelerated computing core has faster access speed and higher memory access bandwidth for the game memory. Most of the program data is in the main memory, and the acceleration thread transfers the data required by the current program segment to the local memory, and writes it back to the main memory after the calculation is completed, freeing up the local memory space for the d...

AI Technical Summary

Problems solved by technology

[0004] Some implicit programming languages ​​such as OpenACC provide data area pragmas to support data reuse between multiple acceleration segments, but mainly for non-distributed device memory shared by all accelerated computing cores such as GPU

On a many-core system with distributed local memory, these pragmas cannot describe the distributed reuse of data in the local memory of multiple computing cores, and can only support this kind of coarse-grained data reuse among multiple acceleration segments, but cannot Supports data reuse between multiple computing modules within the same acceleration segment

Images