[Dune] Bounds checking for matrices and vectors
Elias Pipping
pipping at lavabit.com
Sun Jan 22 18:18:28 CET 2012
Hi,
I'd like to talk about classes like
FieldVector from dune/common/fvector.hh,
FieldMatrix from dune/common/fmatrix.hh
DiagonalMatrix from dune/istl/diagonalmatrix.hh
and the likes for a bit.
My understanding -- until I investigated a bit -- was that those classes
do not perform any bounds checking by default and that defines like
DUNE_FMatrix_WITH_CHECKING and
DUNE_ISTL_WITH_CHECKING
were there just for that(*).
Apparently, it's not that easy. And the fact that some classes contain
specialisations for the case where they only store a single element from
the underlying approximation of a field (e.g. double) is adding to the
mess.
Here's a sample program.
<<SNIP
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include <dune/common/fmatrix.hh>
#include <dune/istl/diagonalmatrix.hh>
int
main()
{
Dune::FieldVector<double,1> vec1;
vec1[0] = 0;
vec1[100] = 29;
Dune::FieldVector<double,3> vec3;
vec3[2] = 0;
vec3[100] = 29;
Dune::FieldMatrix<double,1,1> mat1;
mat1[0][0] = 0;
mat1[100][100] = 29;
Dune::FieldMatrix<double,3,3> mat3;
mat3[2][2] = 0;
mat3[100][100] = 29;
Dune::DiagonalMatrix<double, 1> diag1;
diag1[0] = 0;
diag1[100] = 29;
diag1.diagonal(0) = 0;
diag1.diagonal(100) = 29;
Dune::DiagonalMatrix<double, 3> diag3;
diag3[2] = 0;
diag3[100] = 29;
diag3.diagonal(2) = 0;
diag3.diagonal(100) = 29;
return 0;
}
<<SNAP
The pattern is simple: For each of the aforementioned (three) classes,
two objects are created -- once in three dimensions and once in 1D (to
catch specialised implementations). Each construction is followed by one
or multiple pairs of valid (=0) and invalid (=29) assignments (in that order);
All in all, there are eight invalid assignments in the above; which will
be caught by default (ideal situation: none because bounds checking adds
overhead) and how many are caught if both of the aforementioned defines
are set? The current status appears to be:
* 4/8 always fail
* the aforementioned defines do not affect that number
Here is the body of the above program again, with annotations:
Dune::FieldVector<double,1> vec1;
vec1[0] = 0;
//vec1[100] = 29; // fails an assertion (fvector.hh:242)
Dune::FieldVector<double,3> vec3;
vec3[2] = 0;
vec3[100] = 29;
Dune::FieldMatrix<double,1,1> mat1;
mat1[0][0] = 0;
//mat1[100][100] = 29; // fails an assertion (fmatrix.hh:279)
Dune::FieldMatrix<double,3,3> mat3;
mat3[2][2] = 0;
//mat3[100][100] = 29; // fails an assertion (fmatrix.hh:177)
Dune::DiagonalMatrix<double, 1> diag1;
diag1[0] = 0;
//diag1[100] = 29; // fails an assertion (fmatrix.hh:279)
diag1.diagonal(0) = 0;
diag1.diagonal(100) = 29;
Dune::DiagonalMatrix<double, 3> diag3;
diag3[2] = 0;
diag3[100] = 29;
diag3.diagonal(2) = 0;
diag3.diagonal(100) = 29;
I've attached two patches (one to dune-common and one to dune-istl) that
handle the other four cases with assertions that are guarded by
DUNE_FMatrix_WITH_CHECKING; whether that is the correct define is
questionable.
Whether the four cases whose assertions are not guarded by any defines
(other than the absence of NDEBUG, naturally) should stay that way might
be worth discussing as well.
For the discussion that led me to this investigation, see
http://www.dune-project.org/flyspray/index.php?do=details&task_id=1020
Best regards,
Elias Pipping
(*) The latter is actually used in dune-common even if the name might
suggest otherwise; The former can e.g. be found in
ScaledIdentityMatrix::mv() from dune-istl, too;
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-Add-assertions-to-dm-operator-and-dm-diagonal-for-si.patch
Type: text/x-patch
Size: 1895 bytes
Desc: istl-patch
URL: <https://lists.dune-project.org/pipermail/dune/attachments/20120122/ad0906bb/attachment.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-Add-assertions-to-fvector-vec_access-for-size-1.patch
Type: text/x-patch
Size: 1093 bytes
Desc: common-patch
URL: <https://lists.dune-project.org/pipermail/dune/attachments/20120122/ad0906bb/attachment-0001.bin>
More information about the Dune
mailing list