Unable to render any scene using the hair shape

[amyspark]

All hair-related rendering is not working as of 77d0885.
I've tried Benedikt Bitterli's Fur Ball, Hair Curls and Straight Hair scenes. In all of them the console spams (snippet below):

2017-09-11 08:39:09 WARN  wrk2 [ImageBlock] Invalid sample value : [nan, nan, nan, 1, 1]
2017-09-11 08:39:09 WARN  wrk1 [ImageBlock] Invalid sample value : [nan, nan, nan, 1, 1]
2017-09-11 08:39:09 WARN  wrk2 [vector.h:538] Vector3: Division by zero!
2017-09-11 08:39:09 WARN  wrk0 [vector.h:538] Vector3: Division by zero!
2017-09-11 08:39:09 WARN  wrk3 [vector.h:538] Vector3: Division by zero!
2017-09-11 08:39:09 WARN  wrk0 [MIPMap] evalBilinear(): encountered a NaN!
2017-09-11 08:39:09 WARN  wrk3 [MIPMap] evalBilinear(): encountered a NaN!
2017-09-11 08:39:09 WARN  wrk3 [EnvironmentMap] pdfDirect(): encountered a NaN!
2017-09-11 08:39:09 WARN  wrk1 [vector.h:538] Vector3: Division by zero!
2017-09-11 08:39:09 WARN  wrk0 [EnvironmentMap] pdfDirect(): encountered a NaN!

and ends up with pitch black hair being rendered. The Curly Hair scene instantly crashes.

This happens in MacOS Sierra 10.12.6 as well as Windows 7 (MSVC 2017).

[wjakob]

Dear @amyspark,

we're currently 100% occupied with other research projects. It would be helpful if you could look into this yourself and post a PR to address the underlying issue.

Best,
Wenzel

[amyspark]

I'm no expert, but I'll do my best.
I started by intercepting the division by zeros, which in my case were caused by the integrator attempting to calculate the incident light here:

mitsuba/src/librender/integrator.cpp

Line 183 in 933a057

spec *= Li(sensorRay, rRec);

When Mitsuba intersects the sampled ray with the hair, this query

mitsuba/src/librender/integrator.cpp

Line 150 in 933a057

RadianceQueryRecord rRec(scene, sampler);

generates an Intersection with the shading frame and UV partial position all set to zero. This is the same with any bsdf, and when it tries normalising that vector, it all fails.

[amyspark]

I tried git bisecting since 1ca0413 (when you bumped Mitsuba to 0.5.0). I couldn't get Mitsuba to compile there (lots of Boost uncaught exceptions, clang complaining about libstdc++, et al) so I resorted to manually looking backwards starting from the latest commit.

Counting from today, the first working commit I could find is b8480a3. The first bad one I could find is 8365668.
git bisecting them tells me the bug was introduced by ce80ddc, which is consistent with the shading frame issue I had observed.

EDIT: I've narrowed it down to util.cpp and/or skdtree.h. The changes in hair.cpp only remove the shading frame computation.

[amyspark]

I've finally found the cause of the bug. It seems that previously, Hair's shading frame was initialized to the geometry frame with no further changes:

mitsuba/src/shapes/hair.cpp

Line 848 in 0a49106

its.shFrame = its.geoFrame;

Now, the shading frame calculation is standardized with the rest of the shapes. This was not done anywhere previously:

mitsuba/include/mitsuba/render/skdtree.h

Line 424 in ce80ddc

computeShadingFrame(its.shFrame.n, its.dpdu, its.shFrame);

EDIT: The reason why this all breaks down is because its.uv, its.dpdu and its.dpdv are never initialized at fillIntersectionRecord:

mitsuba/src/shapes/hair.cpp

Line 827 in 933a057

/* No UV coordinates for now */