topic Quote:Some elements in the in Intel® oneAPI Math Kernel Library

different results between mkl_dcscmv and gemv

shaopeng_z_ — Mon, 24 Dec 2012 12:36:59 GMT

In Intel® MKL mkl_dcscmv and gemv produce different result .

At first,I use the mkl_dcscmv to compute K*X=ANS ,where matrix K(2280,2280) converts to the sparse format CSC (COPPTR,ROWIND,VALUES or ROWIND,VALUES,pointer_B,pointer_E) as follows:

program CSC_MATRIX_VECTOR
implicit none
integer::i,j
integer,parameter::nRows=2280

integer,parameter::nNonzero=25395

integer,allocatable::COPPTR(:),ROWIND(:),pointer_B(:),pointer_E(:)
double precision,allocatable::VALUES(:)
double precision,allocatable::X(:),ANS(:)
double precision::alpha,beta
character(6):: matdescra

allocate(COPPTR(nRows+1))
open(unit=10,file='COPPTR.txt')
read(10,*)COPPTR
close(10,status="keep")

allocate(ROWIND(nNonzero))
open(unit=10,file='ROWIND.txt')
read(10,*)ROWIND
close(10,status="keep")

allocate(VALUES(nNonzero))
open(unit=10,file='VALUES.txt')
read(10,*)VALUES
close(10,status="keep")

allocate(X(nRows))
allocate(ANS(nRows))

open(unit=10,file='X.txt')
read(10,*)X
close(10,status="keep")

alpha=1.0
beta=0.0
matdescra='SLNF'

allocate(pointer_B(nRows))
allocate(pointer_E(nRows))

do i=1,nRows
pointer_B(i)=COPPTR(i)
pointer_E(i)=COPPTR(i+1)
end do

call mkl_dcscmv('N', nRows, nRows, alpha, matdescra, VALUES, ROWIND, pointer_B,
pointer_E,X, beta, ANS)

open(unit=10,file='ANS.txt')
write(10,'(D24.16)')ANS
end

result::

ANS(CSC).txt

Secondly,I use gemv to compute K*X=ANS,as follows:

include 'blas.f90'
program FULL_MATRIX_VECTOR
use MKL95_BLAS
use f95_precision
implicit none
integer::i,j
integer,parameter::nRows=2280
double precision,allocatable::K(:,:),X(:),ANS(:)
double precision::alpha,beta

allocate(K(nRows,nRows))
open(unit=10,file='K.txt')
read(10,*)((K(i,j),j=1,nRows),i=1,nRows)
close(10,status="keep")

allocate(X(nRows))
open(unit=10,file='X.txt')
read(10,*)X
close(10,status="keep")

allocate(ANS(nRows))

alpha=1.0
beta=0.0

call gemv(K, X, ANS,alpha ,beta ,'N')

open(unit=10,file='ANS.txt')
write(10,'(D24.16)')ANS
end

the result is

ANS(full).txt

Some elements in the results are quite different,I don't understand why the mkl_dcscmv and gemv cannot produce the same accuracy.

Thanks for your help.

Quote:Some elements in the

mecej4 — Mon, 24 Dec 2012 13:32:12 GMT

Some elements in the results are quite different,I don't understand why

The result vectors that you gave (ansfull.txt,anscsc.txt) have elements as large as 1.6 E4, but the largest difference between the two is 3.7 E-6. On what basis do you say that the results are "quite different"? How much precision do you have in the input matrix and vector, and is a variation of 3.7 E-6 between the two results quite reasonable in that perspective?

I have double precision ,15

shaopeng_z_ — Mon, 24 Dec 2012 14:48:49 GMT

I have double precision ,15 or16 significant digits are accurate. I think the order of magnitude may be 1.0E-10 at least. Why the two methods(mkl_dcscmv and gemv) in the same MKL libary produce different results?

Quote:I have double precision

mecej4 — Mon, 24 Dec 2012 15:27:56 GMT

I have double precision ,15 or16 significant digits are accurate. I think the order of magnitude may be 1.0E-10 at least.

How did you come to this conclusion? Things may work out in that fashion for some matrices, and not so for others. You did not show the input file K.txt, which. presumably, contains the same coefficients as in the compressed matrix, augmented by zero elements. We have to establish that the two K matrices are mathematically identical before we compare the two results for K.x.