Repository navigation
ddot, memory alignment and reproducibility #317
Description
Hi all,
I have noticed that the result of ddot depends on memory alignment of its arguments. Is this expected ? Do some other routines of OpenBlas share the same feature ?
I use xianyi-OpenBLAS-v0.2.4-0-g3a26470.tar.gz with the Intel compiler 13.1.4.183 (I expect it would be the same with gcc, since ddt is written in assembly).
I can provide a test case if it is desired.
Regards,
Philippe Marguinaud
Activity
xianyi commentedon Nov 14, 2013
Hi @pmarguinaud ,
Thank you for the report. It might be a bug.
Please provide the test case.
Xianyi
pmarguinaud commentedon Nov 15, 2013
Hello Zhang,
Here is the code
PROGRAM MAIN
IMPLICIT NONE
DOUBLE PRECISION DDOT
EXTERNAL DDOT
INTEGER, PARAMETER :: N = 5, LDA = 5, J = 5
REAL (8) :: A (LDA_N), AJJ, X (LDA_N*2)
A = (/ &
& 0.1772004544267360204302E+01_8, &
& 0.0000000000000000000000E+00_8, &
& 0.0000000000000000000000E+00_8, &
& 0.0000000000000000000000E+00_8, &
& 0.0000000000000000000000E+00_8, &
& 0.9593654956366169661575E-01_8, &
& 0.8836267311005768787524E+00_8, &
& 0.0000000000000000000000E+00_8, &
& 0.0000000000000000000000E+00_8, &
& 0.0000000000000000000000E+00_8, &
& -0.5078993612458535533705E+00_8, &
& 0.9944539528130179606435E+00_8, &
& 0.1811932610333222948285E+01_8, &
& 0.0000000000000000000000E+00_8, &
& 0.0000000000000000000000E+00_8, &
& 0.9311488401630136912956E+00_8, &
& -0.8367007889097397566047E+00_8, &
& -0.1321798463159503889131E+01_8, &
& 0.1416242430430413312692E+01_8, &
& 0.0000000000000000000000E+00_8, &
& -0.4063195159061600558736E+00_8, &
& 0.3609905431122589214432E+00_8, &
& 0.5710157887908507978736E+00_8, &
& 0.4600399202697713629817E-01_8, &
& 0.1980000019073486328125E+01_8 &
/)
X = 0
X (1:LDA_N) = A
AJJ = DDOT( J-1, X( 1 + LDA * (J-1) ), 1, X( 1 + LDA * (J-1) ), 1 )
WRITE (_, '(E30.22)') AJJ
X = 0
X (2:LDA_N+1) = A
AJJ = DDOT( J-1, X( 1 + LDA * (J-1) + 1 ), 1, X( 1 + LDA * (J-1) + 1 ), 1 )
WRITE (_, '(E30.22)') AJJ
END
I get two slighly different results
0.6235851195535556668048E+00
0.6235851195535555557825E+00
Regards,
Philippe
xianyi commentedon Nov 25, 2013
Hi Philippe,
I checked ddot assembly kernel. This difference is expected.
The dot need sum the x*y result. We want to use SIMD instructions, e.g. aligned load, packed add, packed multiply. However, the different alignment results in the different addition order. Thus, you can get two different results.
Xianyi
pmarguinaud commentedon Nov 26, 2013
Hello Xianyi
Thank you for looking at this. Note, that for the application I am working on, we will have to go back to regular blas/lapack from netlib, since bit-wise reproducibility is a big concern.
Best regards,
Philippe
Date: Mon, 25 Nov 2013 07:08:54 -0800
From: notifications@github.com
To: OpenBLAS@noreply.github.com
CC: pmarguinaud@hotmail.com
Subject: Re: [OpenBLAS] ddot, memory alignment and reproducibility (#317)
Hi Philippe,
I checked ddot assembly kernel. This difference is expected.
The dot need sum the x*y result. We want to use SIMD instructions, e.g. aligned load, packed add, packed multiply. However, the different alignment results in the different addition order. Thus, you can get two different results.
Xianyi
—
Reply to this email directly or view it on GitHub.