[petsc-users] Load Vec from 1D HDF5 dataset MATT READ THIS EMAIL!

[Matthew Knepley]

On Sun, Mar 29, 2015 at 4:00 PM, Håkon Strandenes <haakon at hakostra.net>
wrote:

> Den 2015-03-29 21:48, skrev Matthew Knepley:
>
>>
>> But of course there already is a PETSc-specific HDF5 format. Someone
>> made up "/MESH/nodes/x". This is the essence of HDF5,
>> XML, etc. that they are not actually formats, but rather storage
>> technologies like the UNIX filesystem. Someone still has to declare
>> a format, and we do.
>>
>> However, if you want the data array to be 1D in any old HDF5 tool,
>> except when complex is used, except when timestepping is used,
>> then Yes, we need to make an exception for blocksize. I just think we
>> will end up changing this back when it becomes apparent that
>> having to put a check in all these HDF5 tools as well is onerous.
>>
>>    Matt
>>
>
> My problem, as previously written, is that in the HDF5 output, you do
> already disregard dimensions with size one, all the time, except when
> dealing with the block size. Why?? Just think of the following cases:
>
> 1: When you create a dof=1 DMDA, that dof is not added as an extra
> dimension in the HDF5 output. Why don't you always create this dimension?
> When you create a dof=2 or dof=3, you suddenly write one extra dimension.
>

Yep, I don't like this either


> 2: When you create a 2D DMDA (or 1D), you do not have that third dimension
> in the output. Why isn't a 2D grid just a special case of 3D with one
> dimension where the size is one?
>

True


> 3: When you do not use complex numbers, you do not have that extra
> dimension in the HDF5 output. Why don't you always add one extra dimension
> here in order to be consistent?
>

I really hate this, but type handling is so weak in C and HDF5.


> 4: When you do not use timestepping, this temporal dimension is not added
> in the HDF5 output. Why don't you always add one extra dimension for time,
> just to avoid having to check for it? You coud treat "no timestepping" as
> "timestep 1"?
>

Yep.


> So the ONLY TIME you always add one dimension of size 1 is when dealing
> with the block size. In more or less all other cases you disregard these
> one-size dimensions and tries to compact the dataset into as few dimensions
> as possible. This is HIGHLY non-consistent and non-intuitive.
>
> Is it really that hard to handle this "last special case", and suddenly
> get a consistent HDF5 output format?
>

Okay, I am convinced by the timestepping and dimension. Unstructured grids
turn out to be much more beautiful
since there is none of this messing around.

I guess we should document this somewhere since the PETSc HDF5 format is
not completely obvious.

  Thanks,

     Matt


> Regards,
> Håkon
>
> (I know that I can make a 3D DMDA with one or two dimensions of size 1,
> and then have one or two dimension of size 1 in the HDF5 output, but that
> is really the only case I can think of. This is however not an issue, since
> it is explicitly declared a 3D case, and the user then expect 3D output)
>



-- 
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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20150329/edbae8d3/attachment-0001.html>