MATSHELL a bit broken...
bsmith at mcs.anl.gov
Thu Oct 25 16:41:42 CDT 2007
It should be handled more systematically, inside Mat_Shell
have an entire _MatOps table, also have a PetscBT() with one bit for
each operation. For methods that cannot be overloaded directly turn
that bit on. Now MatShellSetOperation() will check the corresponding bit
if it is set then put the provided function into the _MatOps table that
is inside the Mat_Shell. If the bit is not set then put the provided
function into the regular _MatOps
Make sense? (I don't know that is why I am asking everyone).
Overtime we will likely find that more methods must have the bit set.
On Thu, 25 Oct 2007, Lisandro Dalcin wrote:
> It seems MATSHELL implementation is a bit broken. If I'm not wrong, if
> a user set the operations MOP_ASSEMBLY_END, then the use of the
> automatic MatScale/MatShift would lead to unexpected result...
> Should this operation be managed the same way as
> mult/multtranspose/get diagonal currently is, that is, by backupping
> the user provided function pointer?
> BTW, sorry if I disturb you asking for making changes that seems
> trivial, but I really do not like to change any line of code after
> knowing your opinion.
More information about the petsc-dev