readme_matrixbymatrix.txt

adsp2116x的矩阵乘运算

****************************************************************************

matrixbyvector.asm      Multiplication of a (MxN) matrix by a (NxP) matrix

Analog Devices, Inc.
DSP Division
Three Technology Way
P.O. Box 9106
Norwood, MA 02062

14-JUN-2000 GGL

This directory contains an example ADSP-21160, single-channel, SIMD subroutine 
that implements a matrix multiplied by a matrix; (MxN)*(NxP). 

Files contained in this directory:

matrixbymatrix.dpj  VisualDSP project file
matrixbymatrix.asm  ADSP-21160 source for multiplying a matrix by a vector
matrix_test.asm     Calling function for matrixbymatrix.asm
ADSP-21160.ldf      Linker description file 
mat_a.dat	    Sample data for matrix A
mat_b.dat	    Sample data for vector B
output_exp.dat      Expected output for mat_c where mat_c=mat_a * mat_b
 _________________________________________________________________

CONTENTS

I.    FUNCTION/ALGORITHM DESCRIPTION
II.   IMPLEMENTATION DESCRIPTION
         

I. FUNCTION/ALGORITHM DESCRIPTION

The project matrixbymatrix.dpj contains an implementation of a single-channel, SIMD subroutine that implements a matrix multiplied by a vector.  This routine has been optimized to take advantage of the repetitive multiplies and adds inherent to the algorithm for multiplying a matrix by a vector.  

II.   IMPLEMENTATION DESCRIPTION

Half of the multiply accumulates are performed in PEx and the other half are performed in PEy.  

Please take note that mat_a and mat_b are stored in row major order and mat_c is stored in column major order.  See section 3.1 of the ADSP-21000 Family Applications handbook entitled "Storing a Matrix" for more details on the matrix storage method.  

There is further information on input files and code implementation in the source code comments.