Zero UV partials for custom BSDF

[msqrt]

Hi! I'm trying to implement a custom SVBRDF model with PyTorch -- everything else seems to be working, but the duv_dx and duv_dy fields of the surface interaction are always all zeros. I've added the needs_differentials function that returns True, but it never gets called. This seems to be caused by this code. has_uv_partials for the surface interaction seems to be true every time, but I'm not sure what the dr::is_diff_v<Float> is, and the comment makes me worried that this is just not supported at all.

Is there something I should configure differently to make this work? Or should I compile Mitsuba myself and try to fix the issue in the linked code?

[njroussel]

Hi @msqrt

Indeed, it seems that this path of code is only enabled in non-AD variants. I wonder why...

You can either re-compile the project yourself and remove that constexpr ... is_diff_v<Float condition. Another solution would be to explicitly call si.compute_uv_partials right before calling si.bsdf(ray) in your code.

[msqrt]

Alright, thanks! I changed that and compiled a custom version and it works now -- I ended up having to change the path tracer as well, since it doesn't propagate the ray differentials even for the first bounce.