Re: Revising numeric overflows document

• To: netcdf-hdf@xxxxxxxxxxxxxxxx
• Subject: Re: Revising numeric overflows document
• From: Quincey Koziol <koziol@xxxxxxxxxxxxx>
• Date: Tue, 13 Apr 2004 21:37:20 -0500 (CDT)

Hi Ed,

> > > Quincey Koziol <koziol@xxxxxxxxxxxxx> writes:
> > > 
> > > > Hi all,
> > > >     I've posted a short document describing the changes I think are 
> > > > necessary
> > > > to accomodate supporting netCDF-3 datatype conversions here:
> > > >     http://hdf.ncsa.uiuc.edu/HDF5/planning/DP/Overflow.html
> > > > 
> > > >     Please comment and send suggestions to me.  If there's enough 
> > > > interest, we
> > > > can have a separate meeting to discuss the changes.
> > > > 
> > > >     Quincey
> > > 
> > > Let me ask a question...
> > > 
> > > I would first define a dataset and set it's fill value.
> >     Yes.
> > 
> > > Then I would pass a pointer to my version of the conversion function?
> > > What would it do? Just copy the fill_buf to dst_buf?
> >     You would then need to create a dataset transfer property list and set
> > the conversion function to your callback routine.
> 
> OK, that's cool.
> 
> > 
> > > Then I would try to transfer data.
> >     Yes, using the dataset tranfer property list with your callback routine 
> > set
> > as the conversion function.  Then, during the transfer of the data, if any
> > exceptions occur, your callback routine would be invoked and could handle
> > substituting the fill value for the destination value as desired.
> 
> OK, let me think about this tonight after my kids go to bed and I'll
> see if I understand what is going on.
> > 
> > > Would I know that the callback function has been called? 
> >     It's an "event driven" architecture, so you wouldn't necessarily know it
> > had been invoked, unless you put in code to track that.
> 
> I need to give a range error when the fill value substitution takes
> place. 
> 
> For example, if someone calls nc_put_var_float(), with a bunch of
> doubles, and a fill value has to be used because some are out of
> range, then nc_put_var_float needs to return NC_ERANGE instead of
> NC_NOERR.
> 
> So I have to be able to detect that one or more fill value
> substitutions has taken place. I could set some kind of global
> variable, but that's a bit cheesy, I'm sure you agree.
> 
> It would be very handy indeed if I could somehow get back some
> information about whether any overflows occured...
    Ah!  You could use the "user data" field to pass in a pointer to whatever
you'd like.  Then, you can use that to track any overflows that occur, etc.
Will that work?

    Quincey