A matrix transposition method and system based on a Shenwei 26010 processor

Abstract

The invention relates to a matrix transposition method based on a Shenwei 26010 processor, comprising the following steps of S1 dividing a matrix A stored in a main core into 64 sub-matrices, and numbering the 64 sub-matrices; S2 numbering the 64 slave cores corresponding to the numbers of the 64 submatrices, and respectively reading the 64 submatrices into the slave cores corresponding to the submatrices; S3 transposing submatrices in each slave kernel to obtain 64 transposed slave kernels; S4 arranging 64 transposed slave cores into matrix B in the form of 8*8 according to the numbering order of the slave cores, and transposing the matrix B through inter-core register communication to obtain matrix C; S5 storing the matrix C in the main core, and completing the transposition. By decomposing the larger matrix into smaller blocks, the transposing efficiency is improved by transposing the block matrix in parallel with the matrix transmission.

Description

technical field [0001] The invention relates to the field of matrix transposition methods for processors, in particular to a matrix transposition method and system based on the Shenwei 26010 processor. Background technique [0002] The Shenwei 26010 processor is a high-performance computing processor independently developed by my country. The processor uses the extended ALPHA architecture instruction set, a processor uses 4 core groups, each core group consists of an operation control core (main core, 64-bit RISC structure general processor unit) and an operation core array, namely 8 It consists of 64 computing cores (slave cores) with a mesh structure of 8 columns. Both the main core and the slave core support 256-bit vector floating-point instruction extension; each slave core contains 32 registers, 64KB user-controllable LDM and 16KB program space, and the delay of direct access to the local LDM is extremely small, and the slave core hardware pipeline supports Simultane...

AI Technical Summary

Problems solved by technology

After the calculation is completed, if matrix transposition is required, the data needs to be accessed through direct memory access DMA, and the data is sent back to the master core for data transposition. The DMA sends the data to the slave core LDM, the operation process is very cumbersome, which greatly limits its operation performance

Images