dolfin team mailing list archive

Thread
Date

Cleaned up new GMRES solver. Since PETSc 2.2.1 uses a different interface [...]

To: dolfin-dev@xxxxxxxxxx
From: dolfin@xxxxxxxxxx
Date: Sun, 13 Feb 2005 00:03:07 +0100
Reply-to: Discussion of DOLFIN development <dolfin-dev@xxxxxxxxxx>
Sender: Anders Logg <logg@xxxxxxxxxx>

Commit from logg (2005-02-13 00:03 CET)
----------------

Cleaned up new GMRES solver. Since PETSc 2.2.1 uses a different interface
for KSP and 2.2.0 is not available for Debian, we need to create a new
option for configure, something like --with-petsc-version=2.2.1 in
combination with some #ifdefs in NewGMRES. I won't have time to fix this
until next week, so feel free to take a look.

There's still a hard-to-find memory allocation bug when using the
new GMRES solver in combination with a VirtualMatrix (matrix-free).
Repeated solves (see src/demo/solvers/ode/homotopy/simple) gives
the following error message from BLAS:

** On entry to DGEMV  parameter number  6 had an illegal value

The error does not show up when the program is run through valgrind
(a memory allocation debugger), and valgrind reports no errors...
Anyone have an idea?

  dolfin  src/demo/solvers/ode/homotopy/simple/main.cpp  1.7
  dolfin  src/kernel/la/NewGMRES.cpp                     1.13
  dolfin  src/kernel/la/NewMatrix.cpp                    1.13
  dolfin  src/kernel/la/VirtualMatrix.cpp                1.8
  dolfin  src/kernel/la/dolfin/NewMatrix.h               1.11
  dolfin  src/kernel/ode/MonoAdaptiveNewtonSolver.cpp    1.16
  dolfin  src/kernel/ode/dolfin/ODE.h                    1.32