Fix reading of vector attributes with only one contained value

[franzpoeschel]

Noticed by @PrometheusPi
Some of our backends can't distinguish attributes of scalar types and attributes of vector types containing only a single value
We already check for that situation in some places, in others we forget it.

We will keep forgetting it, so this PR modifies the Attribute::get() function template to do some more automatic conversions:

• Convert from scalar values to vectors with a single value
• Convert between std::vector and std::array types (some backends cannot distinguish those either, so I'll take the chance to cover that)

First commit has a failing test, I'll push the fix afterwards. I will also take the freedom to do some simplifications in Attribute.hpp.

TODO:

• Remove those manual checks that are scattered across our code base?

[ax3l]

Thanks, good idea!

Remove those manual checks that are scattered across our code base?

👍

CI: note that we see a double free in one of the tests currently.

[franzpoeschel]

CI: note that we see a double free in one of the tests currently.

It looks like icpc doesn't really understand variantSrc::visit. I've gone back to the old way for Intel compilers again.

[franzpoeschel]

Remove those manual checks that are scattered across our code base?

Done that now, but I'm leaving this one in to stay on the safe side and always read unitDimension to its dedicated type.

[ax3l]

I think that's a good approach and the PR looks great, thanks a lot!
I'll just slightly modify the ICC guard and then it can go in.

[ax3l]

@franzpoeschel given the original issue that @PrometheusPi saw: what backend and attribute exactly was that?
Was that with PIConGPU?

[PrometheusPi]

@ax3l This is triggered by 0.6.0-dev PIConGPU data created with openPMD-api using the adios2 backend. The error appeared when using the latest openPMD-viewer (see also discussion on SLACK):

In [3]: ts=OpenPMDTimeSeries('/__workdir__/jet_preexpansion_502/simOutput/openPMD/',backend='openpmd-api')
RuntimeError                              Traceback (most recent call last)

<ipython-input-3-b2fd3d9a7bec> in <module>
----> 1 ts=OpenPMDTimeSeries('/__workdir__/jet_preexpansion_502/simOutput/openPMD/',backend='openpmd-api')

~/.conda/envs/analysis/lib/python3.9/site-packages/openpmd_viewer/openpmd_timeseries/main.py in init(self, path_to_dir, check_all_files, backend)
     85 
     86         # - Extract parameters from the first file
---> 87         t, params0 = self.data_reader.read_openPMD_params(self.iterations[0])
     88         self.t[0] = t
     89         self.extensions = params0['extensions']

~/.conda/envs/analysis/lib/python3.9/site-packages/openpmd_viewer/openpmd_timeseries/data_reader/data_reader.py in read_openPMD_params(self, iteration, extract_parameters)
    135 
    136         elif self.backend == 'openpmd-api':
--> 137             return io_reader.read_openPMD_params(
    138                     self.series, iteration, extract_parameters)
    139 

~/.conda/envs/analysis/lib/python3.9/site-packages/openpmd_viewer/openpmd_timeseries/data_reader/io_reader/params_reader.py in read_openPMD_params(series, iteration, extract_parameters)
     75             metadata = {}
     76             metadata['geometry'] = field.get_attribute('geometry')
---> 77             metadata['axis_labels'] = field.axis_labels
     78 
     79             # Swap the order of the labels if the code that wrote the HDF5 file

RuntimeError: getCast: no cast possible.

Most likely caused by reading /data/0/fields/picongpu_idProvider/axisLabels but @franzpoeschel knows this better.

[ax3l]

Clarified offline:

• seen with idProvider (1d mesh)
• axisLabels is only entry then
• ADIOS1/2 attributes are always 1D arrays (of size 1-N)