A weird(?) behavior of shapes_dr

[Mephisto405]

In the below code, I guess snippet (1) and snippet (2) both work exactly the same way, by the definition of ShapePtr, which is not the case here.
More specifically, snippet (1) raises an error DiscreteDistribution: no probability mass found!
Indeed, I intentionally tried to remove these shapes, which raise the error, by the try-except statements. Thus snippet (2) works properly without the error.
Any help would be appreciated!! 🙏🏻🙏🏻🙏🏻

For reproducing the problem:

import drjit as dr
import mitsuba as mi
import torch

mi.set_variant("cuda_ad_rgb")

"""
wget https://benedikt-bitterli.me/resources/mitsuba/staircase.zip -P data/
unzip "data/*.zip" -d data/
rm data/*.zip
"""
scene = mi.load_file("/YOURPATH/data/staircase/scene_v3.xml")

shapes = scene.shapes_dr()
idx = []
for i, shape in enumerate(scene.shapes()):
    try:
        ps = shape.sample_position(0.5, [0.5, 0.5]).p.torch()
        print(i, ps.mean(), shape.surface_area())
        idx.append(i)
    except:
        continue
shapes_reduced = dr.gather(mi.ShapePtr, scene.shapes_dr(), idx)

"""Snippet (1) This does not work"""
shapes_reduced.sample_position(0.5, [0.5, 0.5])

"""Snippet (2) This does work"""
# for i, j in zip(range(len(shapes_reduced)), idx):
#     print(j)
#     shapes_reduced[i].sample_position(0.5, [0.5, 0.5])

[njroussel]

Hi @Mephisto405

Basically shapes.reduces.sample_position is a virtual function call. Unless you disable vcall recording with dr.set_flag(dr.JitFlag.VCallRecord, False), Dr.Jit will "trace/read" through every possible target function and then assemble a single optimized kernel which is finally executed.

In your case, even though some target function will not actually be executed, it is still recorded and during the recording an error happens. I've looked at the error myself, it seems that some of the meshes in the scene are "broken". For example Mesh432.obj is just a line of vertices.

You have 3 options:

• Do a non-parallel evaluation as you suggest in snippet 2
• Disable virtual function recording (locally around the scope of your sample_position call)
• Fix the scene such that it doesn't have any broken meshes