Python Path Integrator

[andiac]

Hello! I am trying to replicate the path integrator in Python, here is my code.

I notice that Mitsuba 3 utilizes the dr::loop, may I know how to replicate it in Python?

I tried to replicate the C++ path integrator line by line, but the output of the Python version is still brighter than the C++ version. Could you please give me an example that the path integrator is written in Python?

[ziyi-zhang]

In Python, the recorded loop will be

loop = mi.Loop("name", lambda: (var1, var2 ... ))
while loop(condition):
    body...

You can find more examples in test_loop.py.

For a Python example of the path integrator, it could be helpful to take a look at prb.py. If you ignore the logic in the if not primal: block, that's essentially a path integrator with standard MIS. A tiny difference is that the path-replay version inherits from the RBIntegrator, while the conventional path integrator can be built on an ADIntegrator.

[andiac]

Thanks!

[andiac]

Hi @ziyi-zhang ! I am reading prb.py and get a question: In line 135, the variable active_next is defined to be

active_next = (depth + 1 < self.max_depth) & si.is_valid()

However, this does not utilize the active calculated in the previous step. The following line seems more reasonable to me:

active_next = active & (depth + 1 < self.max_depth) & si.is_valid()

May I know why the active in the previous step is ignored here? Currently I guess it might be the mechanism of DrJIT's loop.

[Speierers]

The active mask will already be accounted for in while loop(active) hence it isn't necessary to add it here. In other words, the loop condition behaves like a mask to every operation occuring in the loop body. This will be better explained in the upcoming Dr.JIT documentation.

[andiac]

@Speierers Thank you very much for the explanation!