init empty instance

[Ginko2501]

Hi,

I was writing my custom renderer and noticed this subtle difference--

If I initialize my mi.SurfaceInteraction3f by si = mi.SurfaceInteraction3f(), I cannot call si.initialize_sh_frame() later even though I have assigned si.sh_frame.n. On the other hand, if I initialize by si = dr.zeros(mi.SurfaceInteraction3f), I will have no problems.

I wonder what are the differences between these two initialization methods, and which one is the standard way to go. Thanks.

[njroussel]

Hi @Ginko2501

The default constructor will just allocate the memory, but will not initialize it.
For initialize_sh_frame() to work, both the si.sh_frame.n and dp_du fields needs to be initialized. This is why it would crash if you only initialize the former.

As a general rule of thumb, I use dr.zeros everytime I allocate a new variable. If you take a look at some of the source code (Python or C++) you'll see that this is quite common.

[Ginko2501]

Hi, thanks a lot for the reply! You mentioned a couple of times that I should refer to the source code, so I wonder if you could point me to some core files in the source code. I am developing in the Python API a custom renderer that requires autodiff. Thanks again!

[njroussel]

The source code for prbvolpath and prb_reparam is some of the most complete Python source code we have and covers a lot of what one can do with Mitsuba/Dr.Jit.

The real "core" sources are in C++, I'd typically recommend looking at the integrator.cpp. It's a good place to view the full rendering process. In addition, common.py holds some of the boilerplate and tricks to make the rendering process easily differentiable.