About "fully" 3D GRMHD simulations

I must admit, the 3D simulations were a bit anticlimactic to visualize. I know the grid is three-dimensional, but in the end, we can only see 2D plots in the xy and xz planes! I was expecting something rendered as an actual 3D volume.

With that in mind, I’d like to ask: to generate a 3D volumetric plot like the one in the attached image, is it simply a matter of converting the .hdf5 file into a format readable by ParaView or VisIt? Or do I need to rely on a different tool or a different version of the code (perhaps a GPU-compiled version of HARM)?

The attached image is from an old paper ( https://academic.oup.com/mnrasl/article/394/1/L126/1082160?login=false ), in which they seems to have used HARM for all the simulations.

[c-prather]

3D volumetric rendering was the first thing I wanted too, but it's really much less useful and much more painful to do than you'd think.

The primary problem is that GRMHD codes use an arbitrary (affine) coordinate transformation. None of the major packages (VisIt, Paraview) support "regular" grids which are only regular in weird, bespoke coordinates, so we end up using their facilities for totally unstructured grids, which are extremely slow since nearest cells must be searched for manually. Since our output format uses just a single mesh block, the computation can't even be parallelized with the existing packages (which only parallelize over meshblocks, at least last I checked), and being unstructured no one's likely to get it going fast on a GPU. Due to the constraints, high-res volumetric plots like the above take O(hours) on a single core. That makes 3D plots/movies way too inconvenient to do regularly, so the habit (at least in the parts of the community I'm familiar with) is to make one-off "hero" plots like the above (which is in VisIt, I think?) and then throw away the ugly one-off scripts hoping that by the next time you need them, someone's invented a better way.

What's worse, playing much with volumetric rendering especially, I was quickly convinced you could tune the opacity curve and density isosurfaces to tell just about any story you wanted to tell. There's no diagnostic potential because the colors aren't precise due to lighting and opacity effects, perspective makes it difficult to directly compare shapes, and the turbulence and other fine details are just completely lost. So, I also pretty quickly got bored, and I imagine others did too by the lack of 3D renders in a lot of modern papers.

That said, pyharm has a script write-xdmf which generates XDMF description files intended for VisIt which should allow plotting most variables in a .h5 dump file (either from iharm3d or converted from KHARMA). It will also need access to all the geometry (again, it thinks the grid is "unstructured") which can be generated with pyharm write-grid.
The result is slow and unwieldy, any calculated values like the magnetization are difficult to implement or suspect due to VisIt not being GR-aware, and every time I've dealt with it I've felt like I'm wasting my time. So, there's your warning I guess. If you can implement a better way to plot GR in 3D you would probably have a very well-cited paper out of it.