Marian Piatkowski authored 10 years ago and Steffen Müthing committed 10 years ago

added the old constructors called with the variable recalc_defect to RestartedGMResSolver and set them as deprecated, since they should not be used anymore

Marian Piatkowski authored 11 years ago and Steffen Müthing committed 10 years ago

replaced the usual arrays for the Givens-rotation coefficients and the search directions etc. with the arrays from Dune

implemented the BLAS-routine drotg for greater robustness which was a TODO note to compute the Givens rotation

Marian Piatkowski authored 11 years ago and Steffen Müthing committed 10 years ago

removed bool recalc_defect
reimplement the update function with respect to optimality
fixed strange behaviour when maxiter is reached and GMRes wants to restart infinitely many times
fixed strange behaviour when linear reduction is reached but GMRes wants to restart and continue calculation
fixed the printing of the results

This caused build errors with CMake.

fix "unused variable" warnings by proper use of #ifdef

MatrixAdapter uses a const reference to store a reference
to the matrix it is working on. Without this patch it was
possible to pass the matrix to constructor actually taking
a MatrixAdapter as its argument. This resulted in a const
reference to a temporary MatrixAdapter that died after the
constructor call.

This patch request explicitly calling the constructor and
thus fixes this issue.

Elias Pipping authored 11 years ago and Markus Blatt committed 11 years ago

This problem was introduced in 3da29f78

The last patch missed one instance of double.
# Please enter the commit message for your changes. Lines starting
# with '#' will be ignored, and an empty message aborts the commit.
# On branch master
# Changes to be committed:
#   (use "git reset HEAD <file>..." to unstage)
#
#	modified:   ../../../../dune/istl/test/superlutest.cc
#
# Untracked files:
#   (use "git add <file>..." to include in what will be committed)
#
#	../../../../build-clang/
#	../../../../build-test/
#	../../../../parallel/
#	../../../../test-shared/

Somehow I missed this instance in the transition.

Previously this code was plain wrong. It actually made a copy
of the vector and tried to swap its entries with the original vector.
The argument to swap even was a const reference!
Now we use the correct trick: vector().swap(other).

use FieldTraits to determin the return type of the norm methods

use FieldTraits to determin the return type of the norm methods

update matrixtest to use the FieldTraits and not assume double as the return value of *_norm

use FieldTraits to determin the return type of the norm methods

use FieldTraits to determin the return type of the norm methods

add header ftraits.hh as we are using the FieldTraits now

add blockvector test for nested block vectors

add FieldTraits specialization for BlockVector (derived from the patch by Bernd Flemisch)

fix usage of FieldTraits (similar to the PromotionTraits)

Bård Skaflestad authored 11 years ago and Markus Blatt committed 11 years ago

Inspired by commit fcff302d.

Arne Morten Kvarving authored 11 years ago and Markus Blatt committed 11 years ago

Previously memory was only releases in the case that the data
was copied. This resulted in a memory leak if no copying was
requested. This patch fixes this by moving the deallocation to
the correct scope and consequently closes flyspary #1326, see
http://www.dune-project.org/flyspray/index.php?do=details&task_id=1326

The patches that made copies of AMG usable form different threads
changed its internal interface. This patch adapts KAMG to these
changes and makes it compile and run again. While at it, I changed
from normal pointers to shared_ptr where it seemed suitable.

Jurgis Pods authored 11 years ago and Markus Blatt committed 11 years ago

This is essentially equivalent to Uli Sack's patch from Flyspray task FS#1202, ported to the current master.

Currently, GMRes is inconsistent when comparing current ('norm') and initial defect ('norm_0'). While the current residual is taken from the preconditioned residual, the initial defect - at least in the case the flag '_recalc_defect' is not set, which is the default - is taken from the unpreconditioned residual.

This causes the following problems:
1) The calculated reduction is wrong.
2) The solver might even exit in iteration 0 (without entering the iteration loop) because 'norm' and 'norm_0' differ by a factor greater than the requested reduction, returning a zero update 'x'.

By using the preconditioned residual in any case, the defect calculation is made consistent again.

To make future patches more readable the list of headers is now ordered
alphabetically with one header per line.

Storing the iterators over over the AMG hierarchies as class member
is a really bad idea if one wants to allow the reuse of theses AMG
hierarchies in different threads. This patch creates a private struct
for collecting the iterators over the hierachies. An instance of this
struct is created in apply and passed to the solution algorithms. Thus
each apply uses its own context of iterators.