[Soc-2018-dev] Weekly Report #12 - Many Light Sampling

[Erik Englesson]

Hi all,

*This week* I have been trying my branch on some of the demo scenes and
found bugs that I have started to fix.

In the Barcelona Pavillion scene, I noticed that instancing of mesh lights
did not work. This was fixed in this commit
<https://developer.blender.org/rB84fec2152261>. In the same commit, I also
fixed the emission to luminance conversion to use linear_rgb_to_gray()
instead of my own. I also noted that mesh lights are double sided which I
have not had time to do anything about(considered single sided right now).

When debugging the fishy cat scene that Brecht mentioned last week I found
a bug in the energy calculations for each light source. If it did not have
constant emission then I considered it to have no energy at all. This has
now been changed so it uses an emission of (1,1,1) instead.
The light picking method uses a diffuse approximation term in the
importance metric which depends on the surface normal at the point. This
normal used to make sure that only lights in the same hemisphere as the
normal is pointing towards got sampled. However, this should not be the
case for any surface where refraction is possible. This case came to light
from the transparent cornea of the cat. Both the energy and the normal
issues have been fixed here
<https://developer.blender.org/rB108594d8c844edb09afdac1a355f0b6d99cb80db>.

The classroom scene is a very complex test scene. The window is a mesh
light and a portal that has a normal pointing out from the room, there is a
strong fill light just outside pointing away from the room, the blackboard
lamp consists of a mesh light(non-constant emission) and an area light in
front of it, there is a strong point light in the corridor that is not
visible for most points and finally a strong distant light.
In this scene, I found that the bounding boxes I calculated for area lights
were wrong. This was fixed here
<https://developer.blender.org/rB5d344f43603ddba8f3a009915b92a5374066efeb>.
There was a dustbin that was transparent which made me realize(with the
help of Brecht) that the position and normal that is used for the light
picking should be from the last non-transparent bounce. This change can be
found here <https://developer.blender.org/rBe8e0669785cd>.

Brecht mentioned that I use doubles in the split heuristic to avoid some
precision and overflow issues but doubles will not work well for the GPU. I
have now changed this to use floats instead but I am still using doubles to
calculate the energy variance on the host which seems to be where the
precision issues came from. The overflow problem was related to squaring
the number of emitters in the node. When I saw that it overflowed I think I
used integers to calculate the square which has a much lower maximum value
than floats. It does not seem to overflow anymore at least. This change can
be found here
<https://developer.blender.org/rBf6305047f44bf3168b1da8601307d78015ad8f65>.

I have also continued the work on volumes. I think I have been able to only
use the new light picking strategy for shading points that are not inside
or on the boundary of a volume. For volume points, it is using the old
picking method. This is until the volume part of the paper has been
implemented. I also fixed a bug in light_background_sample() that used an
index into the lights array as an index into the distribution array. The
volume changes and the bug fix can be found here
<https://developer.blender.org/rB94af4326e3fc22fe1b229a3e2435fb6ccfe36ce9>.

*Next week* is a bit unclear to me. Am I only supposed to write
documentation and do the evaluation? If so, what kind of documentation do
you want? There are still things I have not had time to fix, e.g. the clock
in the classroom scene is darker using the light tree, branched vs normal
path tracing has some brightness differences(I have not implemented
splitting in light_bvh_pdf() ), doing something about mesh lights being
double sided, add more comments to the code, splitting is not implemented
for branched path tracing of volumes, add tests, try running on GPU, etc.
Can I work on some of this next week? I would appreciate if someone could
clarify this for me.

Have a great weekend!

Thanks,
Erik
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.blender.org/pipermail/soc-2018-dev/attachments/20180803/affb91f0/attachment.html>