High-performance realization method of gemm dense matrix multiplication on Shenwei 26010 many-core CPU

A technology of dense matrix and implementation method, applied in the fields of instrumentation, calculation, electrical digital data processing, etc., can solve the problem of low performance, open source BLAS math library Shenwei many-core processor 26010 optimized, can not give full play to many-core computing capabilities, etc. problem, to achieve the effect of improving the performance of the function

Active Publication Date: 2020-06-09
INST OF SOFTWARE - CHINESE ACAD OF SCI +1
View PDF6 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0010] The problem solved by the present invention is: based on the fact that the existing open-source BLAS math library has not been specifically optimized for the Shenwei many-core processor 26010, the computing power of the many-cores cannot be fully utilized, and the performance is low. The present invention proposes the Shenwei 26010 A high-performance implementation method of GEMM dense matrix multiplication on many-core CPUs. This method is based on platform features such as storage structure, memory access methods, hardware pipelines, and register-level communication mechanisms on the Shenwei many-core processor 26010 single-core group. It uses data storage formats. , vectorization, calculation and memory access overlapping, loop unrolling and software pipeline and other technologies, apply and improve the block matrix multiplication algorithm, and optimize function performance efficiently

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • High-performance realization method of gemm dense matrix multiplication on Shenwei 26010 many-core CPU
  • High-performance realization method of gemm dense matrix multiplication on Shenwei 26010 many-core CPU
  • High-performance realization method of gemm dense matrix multiplication on Shenwei 26010 many-core CPU

Examples

Experimental program
Comparison scheme
Effect test

Embodiment Construction

[0036] The present invention will be described in detail below in conjunction with the accompanying drawings and embodiments.

[0037] Such as figure 1 As shown, the specific implementation adopts the three-level code framework of "interface interface layer-scheduling task scheduling layer-kernel assembly computing layer", which is described as follows:

[0038] (1) Interface interface layer function: This layer is a function interface, which checks the input parameters and returns an error code if an illegal parameter is judged; in addition, according to the accuracy and transposition of the input matrix A, B, call the corresponding scheduling Task scheduling layer function;

[0039] (2) Scheduling task scheduling layer function: It is called by the interface interface layer function, and calls the kernel assembly computing layer function. The updating order of the matrix C is controlled through the n-k-m three-layer loop, and the C sub-matrices are updated serially, with ...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

PUM

No PUM Login to view more

Abstract

The present invention discloses a high-performance implementation method of GEMM dense matrix multiplication on Shenwei 26010 multi-core CPU, aiming at Shenwei multi-core processor 26010, and based on platform characteristics such as storage structure, memory access, hardware pipeline and register-level communication mechanism, the matrix is ​​optimized Blocking and inter-core data mapping method, designed a top-down three-level block-parallel block matrix multiplication algorithm, designed a slave-core computing resource data sharing method based on a register-level communication mechanism, and used asynchronous DMA data transmission between master-slave cores Mechanism, designed a double-buffering strategy with overlapped calculation and memory access, designed a loop unrolling strategy and a software pipeline arrangement method on a single slave core, used efficient register block mode and SIMD vectorized multiply-add instructions, and implemented functions optimization. Compared with the single-core open source BLAS math library GotoBLAS, this high-performance GEMM function has an average speedup of 227.94 and a maximum speedup of 296.93.

Description

technical field [0001] The present invention relates to a high-performance implementation method of the basic linear algebra library BLAS (Basic Linear Algebra Subprograms) level3 function dense matrix multiplication operation GEMM (General matrix-matrix multiplication), especially the high-performance GEMM implementation based on the Shenwei many-core platform architecture and Optimization. Background technique [0002] As the underlying and basic matrix operation, dense matrix multiplication is widely used in the fields of scientific and engineering computing, so its performance is extremely demanding. BLAS is a basic linear algebra subroutine library, which mainly includes the basic operations of vectors and matrices. It is one of the most basic and important mathematical libraries. At present, most software related to matrix operations will call the BLAS library, and the world's supercomputer TOP500 ranking The bottom layer of the benchmark software LINPACK is supported...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

Application Information

Patent Timeline
no application Login to view more
Patent Type & Authority Patents(China)
IPC IPC(8): G06F9/302G06F9/38
CPCG06F9/3001G06F9/3012G06F9/3814G06F9/3869G06F9/3881G06F9/3887G06F9/3856
Inventor 杨超蒋丽娟尹万旺敖玉龙魏迪袁欣辉刘芳芳张鹏
Owner INST OF SOFTWARE - CHINESE ACAD OF SCI
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap