On Jun 15, 2012, at 11:07 AM, Daniel Lowell wrote:
> I did indeed bypass them.
Never bypass them!
> Ok, thanks.
>
>
> On Fri, Jun 15, 2012 at 11:06 AM, Matthew Knepley <knepley at gmail.com>
> wrote:
> On Sat, Jun 16, 2012 at 12:03 AM, Daniel Lowell <redratio1 at gmail.com>
> wrote:
> Hey Satish,
>
> So I'm having an issue with this function:
>
> #define PetscObjectComposedDataGetReal(obj,id,data,flag)
> \
> ((((obj)->realcomposedstate && ((obj)->realcomposedstate[id] ==
> (obj)->state))? \
> (data = (obj)->realcomposeddata[id],flag = PETSC_TRUE) : (flag =
> PETSC_FALSE)),0)
>
> Called from VecNorm which is called from SNESLineSearchApply_BT
>
> I have two examples running one is the vanilla ex14 using my GPU vector and
> matrix type. The other example is a variant of ex14test where I am
> implementing GPU formfunction and formjacobian.
> ex14 vanilla runs through correctly, however ex14test diverges in the logic
> within VecNorm, specifically line 164:
> /*
> * Cached data?
> */
> if (type!=NORM_1_AND_2) {
> ierr = PetscObjectComposedDataGetReal
> ((PetscObject)x,NormIds[type],*val,flg);CHKERRQ(ierr);
> if (flg) PetscFunctionReturn(0);
> }
>
> The ex14 detects there is a "real" part, however the ex14test does not and
> VecNorm returns.
>
> "val" in VecNorm is PetscReal gnorm = fnorm defined in
> SNESLineSearchApply_BT.
> For ex14: val[0]=3.44e-317, val[1]: 0.580
> For ex14test: val[0]=3.44e-317, val[1]: 0.029
>
>
> This difference makes SNESLineSearchApply_BT branch into different execution
> paths.
>
> I'm not sure where the issue could have come from. Any ideas?
>
> If you change values in a Vec, you must increment the internal state, since
> we cache norms.
> The internal functions for array getting do this automatically, so I am not
> sure why this is
> messing up. Did you bypass them?
>
> Matt
>
>
> --
> Daniel Lowell
>
>
>
>
> --
> What most experimenters take for granted before they begin their experiments
> is infinitely more interesting than any results to which their experiments
> lead.
> -- Norbert Wiener
>
>
>
> --
> Daniel Lowell
>
