How to obtain "throughput" at each surface intersection in mitsuba?

[woAIxuexiSR]

Hi there,

I’m implementing a "path guiding" algorithm in Mitsuba. I’m trying to extract the throughput at each surface intersection during path tracing using the following code:

while loop(active):
    ... # path tracing code
    thp = throughput.torch()
    valid = si.is_valid().torch()

However, this doesn’t seem to produce the correct results. Could you please advise on the proper way to obtain the throughput at each intersection?

Thank you in advance for your help!

[merlinND]

Hello @woAIxuexiSR,

You did not give much information so I am assuming that you are trying to run the latest version in JIT mode (cuda_* or llvm_* variants).
By default, Mitsuba integrators run in symbolic mode, meaning that the path tracing loop gets executed in a single kernel. A great advantage is that intermediate values can be kept in registers, without constantly writing state out to global memory & reading it back in at each bounce.
But it also means that you cannot break out of the loop mid-way to start executing PyTorch code.

You can switch from symbolic mode to wavefront mode with the following flag (I think), but you should expect a big slow-down and increased memory usage:

dr.set_flag(dr.JitFlag.SymbolicLoops, False)

See also this discussion about recording paths: #187