Envmap and constant emitters ignore active mask in pdf_direction()

[Dream4Leo]

I noticed that unlike area and sunsky, envmap and constant ignores the active mask in pdf_direction().

The path integrator calls pdf_emitter_direction() with active = !prev_bsdf_delta. When the previous bounce is a Delta BSDF, this mask is false. I expect to mask the emitter pdf to 0.0 so that the MIS weight for the BSDF strategy resolves to 1.0.

However, constant and envmap return their standard PDF regardless of the mask. Is this the intended behavior?

[merlinND]

Hello @Dream4Leo,

I think that the return values of masked entries (active == false) is undefined, meaning that there is no guarantee either way. The idea is that the value should be unused anyway, its value wouldn't matter so there's no point spending time on it.

In practice though, two things happen:

• Most plugins do mask out their return value, e.g. return result & active;
• Return values of virtual function calls are automatically set to zero in masked lanes

[Dream4Leo]

Thanks for your quick reply. Do you mean that return values of the following function are set to zero in masked lanes, without explicitly writing return warp::square_to_uniform_sphere_pdf(ds.d) & active;?

mitsuba3/src/emitters/constant.cpp

Lines 152 to 157 in 0d082f2

	Float pdf_direction(const Interaction3f &, const DirectionSample3f &ds,
	Mask active) const override {
	MI_MASKED_FUNCTION(ProfilerPhase::EndpointEvaluate, active);
	return warp::square_to_uniform_sphere_pdf(ds.d);
	}

I agree that generally masked entries shouldn't matter. But for the very specific code here, an undefined or non-zero value of em_pdf seems to introduce bias for delta BSDF (mis_bsdf != 1.0) in masked lanes.

mitsuba3/src/integrators/path.cpp

Lines 210 to 215 in 0d082f2

	if (dr::any_or<true>(!ls.prev_bsdf_delta))
	em_pdf = scene->pdf_emitter_direction(ls.prev_si, ds,
	!ls.prev_bsdf_delta);
	// Compute MIS weight for emitter sample from previous bounce
	Float mis_bsdf = mis_weight(ls.prev_bsdf_pdf, em_pdf);

[merlinND]

Hello @Dream4Leo,

seems to introduce bias

Is this something that you have observed in practice, or more of a risk that you suspect from looking at the code?

My understanding is that there are two cases to take into account and both are covered:

1. Scalar variants: the masked case is protected by an if (with em_pdf initialized to zero above)

mitsuba3/src/integrators/path.cpp

Line 210 in 8e624f2

if (dr::any_or<true>(!ls.prev_bsdf_delta))

2. Vectorized variants: the vectorized virtual function call ds.emitter->pdf_direction() will zero-out masked return values automatically.

mitsuba3/src/render/scene.cpp

Line 395 in 8e624f2

return ds.emitter->pdf_direction(ref, ds, active) * emitter_pmf;

[Dream4Leo]

It was indeed a vectorized-specific suspect. Based on your reply, I wrote a script and could verify that the virtual function indeed zero-out masked lines.
I don't quite get into the template magic there but I think it is enough for now. Thank you for your time!