How does direct emitter visibility (from sensor) work in Path integrator ?

[xacond00]

When there is previous surface interaction ls.prev_si, the code makes perfect sense, as the direction sample contains direction towards the previous intersection point, and thus later calling pdf_direction() correctly computes cos(theta) term, based on the surface normal and incoming direction.

But how on earth does this code work, when ls.prev_si is initialized to dr::zeros() for primary rays generated from the sensor ? The direction vector in ds now points from the world origin towards the emitter intersection. How can then correct cos(theta) be computed, when the direction should be identical with primary ray's world direction and not si.p - ref.p where ref.p = (0,0,0) based on the direction sample constructor ?

if (dr::any_or<true>(si.emitter(scene) != nullptr)) { 
                    // On 0th iteration, the prev_si contains all zeros, meaning ds generates direction towards si.p from world origin ?
                    // But shouldn't the direction be from ray.o towards si.p ???
                    DirectionSample3f ds(scene, si, ls.prev_si);
                    Float em_pdf = 0.f;
                   // Cos theta term is internally computed based on ds.d and si.n
                    if (dr::any_or<true>(!ls.prev_bsdf_delta))
                        em_pdf = scene->pdf_emitter_direction(ls.prev_si, ds, !ls.prev_bsdf_delta);
                    Float mis_bsdf = mis_weight(ls.prev_bsdf_pdf, em_pdf);
                    ls.result = spec_fma(ls.throughput, ds.emitter->eval(si, ls.prev_bsdf_pdf > 0.f) * mis_bsdf, ls.result);
}

So I guess my question is, why the prev_si isn't initialized with primary ray's origin as the intersection point ? It just doesn't make sense, how we can measure radiance without knowing the prior point ?

[njroussel]

Hi @xacond00

The prev_bsdf_delta variable is initialized to true, so the entire em_pdf = scene->pdf_emitter_direction(ls.prev_si, ds, !ls.prev_bsdf_delta); computation is masked (the second argument of pdf_emitter_direction is a mask to indicate which lanes/threads are active). Effectively this means that em_pdf is equal to 0 and doesn't depend on the cosine term you mentioned.

Note that prev_bsdf_pdf is initialized as 1, so the mis_bsdf is equal to 1 even though em_pdf is 0.

[xacond00]

Yes, I noticed that too yesterday. Maybe should test more, before bothering anyone here :P
I made a mistake when I was stripping the sample function to compute primary interaction only, and completely missed the negation in there. The !ls.prev_bsdf_delta is infact all false on the primary interaction, and thus the emitter pdf computation doesn't happen at all.

Anyways, thanks !