[Paraview] Reading C-ordered data from hdf5

[Dominik Szczerba]

John R. Cary wrote:
> Dominik Szczerba wrote:
>> There is no other way of reading C arrays into Fortran ordering than 
>> ordering conversion, but it can be implemented optimally as an 
>> in-place transpose (once already in memory) on the reader side. This 
>> costs time, not storage. I so much hoped this would be a feature of 
>> xdmf2 for example, it is one in my own home grown hdf5 Paraview 
>> reader. --DS
> 
> 
> So I guess what I could do is reorder my data pointer array, then set 
> the pointer
> of vtkDataArray to my pointer?

That is what I suggest to try if you care about optimal performance.

> 
> And I should be able to transpose by just swapping the fortran index
> and C index locations, but then I have to figure out how to do it
> just once.

well yes, you will have to "touch" the original data and need to 
transpose it back if necessary and when not anymore needed by VTK 
(remember that with SetArray it will be referenced). Alternatively, you 
can do inverse on a copy, if storage is no issue.

-- Dominik

> 
> Thanks....John Cary
>>
>> John R. Cary wrote:
>>> Dominik Szczerba wrote:
>>>> You will need SetArray method from the concrete implementation of 
>>>> vtkDataArray. I could not quickly find in the source code any hints 
>>>> about the underlying storage ordering, but it is trivial to find out 
>>>> experimentally. Just post the results please. If it is the same as 
>>>> yours no copying is needed.
>>>>
>>>
>>> Well, I was unable to figure out how to get hdf5 to read the data
>>> in by Fortran order, so I wrote the snippet below.  Sending by
>>> request.
>>>
>>> This is probably slow, so better solutions always welcome.
>>>
>>> John Cary
>>>
>>> // The index tuple is initially zero
>>>    size_t* indices = new size_t[rank];
>>>    for(size_t k=0; k<rank; ++k) indices[k] = 0;
>>> // Step through by global C index
>>>    for(size_t k = 0; k<len; ++k) {
>>> // Accumulate the Fortran index
>>>      size_t indx = indices[rank-1];
>>>      for(size_t j = 2; j<=rank; ++j) indx = indx*dims[rank-j] +
>>>        indices[rank-j];
>>> // Set the value in the VTK array at the Fortran index to the
>>> // value in the C array at the C index
>>>      if (H5Tequal(type, H5T_NATIVE_DOUBLE)) rv->SetTuple(indx,
>>>        &((double*) data)[k]);
>>>      if (H5Tequal(type, H5T_NATIVE_FLOAT)) rv->SetTuple(indx,
>>>        &((float*) data)[k]);
>>> // Update the index tuple
>>>      size_t j = rank;
>>>      do {
>>>        --j;
>>>        ++indices[j];
>>>        if (indices[j] == dims[j]) indices[j] = 0;
>>>        else break;
>>>      } while (j != 0);
>>>    }
>>>
>>>
>>>
>>>>>>
>>>>>> John R. Cary wrote:
>>>>>>> I have arrays written out from C into HDF5 using the HDF5 C
>>>>>>> API, and so in the file the arrays are in C ordering.  I now
>>>>>>> need to read them into VTK.  AFAICT, VTK is Fortran ordering,
>>>>>>> even though it is written in C++.  So that means I need to
>>>>>>> reverse the ordering.  Of course, I can write a loop to do
>>>>>>> this, but is there some standard way that VTK folks do this?
>>>>>>> I don't want to reinvent the wheel.
>>>>>>>
>>>>>>> Sorry for what is likely an elementary question, but I could
>>>>>>> not find it on the FAQ or after googling quite a bit.
>>>>>>>
>>>>>>> Thanks so much......John Cary
>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
> 

-- 
Dominik Szczerba, Ph.D.
Foundation for Research
on Information Technologies in Society
http://www.itis.ethz.ch