| 4bfab301 | 16-Jul-2009 |
Barry Smith <bsmith@mcs.anl.gov> |
new policy for extern declaration and handling of CAPS and _names in Fortran for Fortran kernels, this information is now in a .h file parallel to the .F file that contains the kernel
Hg-commit: 03c
new policy for extern declaration and handling of CAPS and _names in Fortran for Fortran kernels, this information is now in a .h file parallel to the .F file that contains the kernel
Hg-commit: 03c15240f3d12a79aee5a3e0c56d9f10c34c6434
show more ...
|
| 14ca34e6 | 14-Nov-2008 |
Barry Smith <bsmith@mcs.anl.gov> |
From: mstefano@milan.westerngeco.slb.com
Subject: [Fwd: Re: [PETSC #18599] Valgrind problem with MatGetSubMatrix]
Date: November 14, 2008 2:30:37 AM CST
To: petsc-maint@mcs.anl.gov
Cc: pets
From: mstefano@milan.westerngeco.slb.com
Subject: [Fwd: Re: [PETSC #18599] Valgrind problem with MatGetSubMatrix]
Date: November 14, 2008 2:30:37 AM CST
To: petsc-maint@mcs.anl.gov
Cc: petsc-maint@mcs.anl.gov
I think I've found the bug.
Look at src/mat/impls/aij/seq/aij.c, line 1699.
There, variable "i" is declared as a PetscInt.
Then, look at line 1710 or 1712.
The variable "i" is used to get the truth answer from the ISSorted function.
As long as you don't build PETSc with --with-64-bit-indices, all is OK.
But (and this is my case) if you built PETSc with 64 bit indices, the PetscTruth data type is smaller (as byte size) than PetscInt. The result is that only part of the bytes of the "i" variable are initialized by ISSorted.
I strongly suggest to add a new variable to be used exclusively as a PetscTruth flag for the ISSorted calls.
Fixed this bug all over the place
Hg-commit: 79369c761f62f38e25800abc8725233db2a5486a
show more ...
|