[Bf-committers] How about adding boost uBLAS librarytoblender/extern?

[Yves Poissant]

From: "joe" <joeedh at gmail.com>
Sent: Monday, January 26, 2009 11:14 AM


> Blender's core math stuff basically stores vectors and matrices as
> float arrays.  Converting the renderer to use specific datatypes would
> likely be a huge pain, and would need a lot of forethought.

I agree. Althopugh, there is nothing in the examples you provided that can 
only be optimized in C and not in C++.

> What sort of profound refactoring are you talking about, btw?

Well I tend to write too much and if I start elaborating on that it can get 
really lengthy. But here I go anyway.

<preambule>
The "profound" or "fundamental" refactoring I talk about is to do with 
representation of materials, lights and illumination calculations. Right 
now, the render engine is a collection of CG tricks that were developped 
through the years by researchers and that were implemented in Blender in an 
add-hoc way. This way of representing material properties and light 
properties and of calculating a shading value on object surfaces I call that 
the "legacy renderer" or "first generation renderer" way. Those tricks were 
developped by researchers that had little insights into real physics of 
material and lights.

Legacy renderers are more or less the extension of 70's state of ad-hoc 
rendering technology with additional tricks that fitted that old model. 
Scanline converters were developped in 1967 by Wylie et al. Phong shading 
was developped in 1975 by Phong and the Phong surface properties are still 
the basis of legacy rendereers today. Z-buffers were demonstrated around 
1975 by Catmull, Myers  and Watkins. The ray-tracing algorithm was 
demonstrated in 1980 by Whitted. By 1985, all the bases of the legacy 
rendering tricks were invented and formalized in text books such as 
"Fundamentals of Interactive Computer Graphics" by Foley and Van Dam in 
1982, "Procedural Elements for Computer Graphics" by Rogers in 1985 "An 
Introduction to Ray-Tracing" by Glassner in 1989. In 1983, Roy Hall 
published "A Testbed for Realistic Image Synthesis" that essentiially 
provide all the material properties that are still in use in todays legacy 
renderers. The basic material properties that we tweak in Blender are all 
defined in this paper.

Then in 1982, Clark proposed the "Geometry Engine" and this marked the date 
of the cristalization of this rendering approach for all the years to come. 
The Geometry Engine evolved into Silicone Graphics and their hardware 
graphics accelerators with their ligrary that allowed to use this hardware 
IrisGL which eventually produced OpenGL. Then there was this race to produce 
yet more powerfull OpenGL acceclerators until today. Direct3D is just 
another proprietary set of API to use OpenGL type of shading.

On the other hand, Cook & Torrance already published some BRDF related paper 
in 1981 but their shading equations were eventually recuperated by the 
legacy renderers and the true result of their research stayed unnotice for 
years. In 1986, Kajiya published "The Rendering Equation" and demonstrated 
the first GI algorithm. But the hardware required to do that was out of 
reach even for some research teams. 1984 to 1986 were the years where 
radiosity algorithms were developped. Then a long period if stagnation 
happened until 1992 and then 1995 to 1997 where efficient GI algorithms were 
finally developped such as the Photon Mapping by Jensen in 1996, the 
Metropolis Light Transpôrt by Veach in 1997 and the Instant Radiosity by 
Keller in 1997. But by that time, OpenGL was the new standard and was 
undetronable.

In 2009, things are changing. Single core computers are out of the way and 
multicore computer are in. It is now possible to do realtime or very near 
realtime GI of very good quality.
</preambule>

Legacy renderers are outdated and need to be replaced by more modern 
renderers. It is impossible to get good surface renders with legacy material 
properties and single sample illumination. It takes considerable time to 
tweak legacy material properties to get a realistic render and it takes even 
more time to do that for animation because legacy material property tweaking 
are view dependent. The modern way to describe materials are based on the 
BRDF concept. The shading algorithm needs to be totaly GI integrated. GI 
meaning Global Illumination, it means that the whole rendering equation 
should be integrated in one renderer instead of adding legacy CG trick 
results. New renderers such as LuxRender, Indigo, YafAray, Maxwell, 
FryRender, etc, and to some extent VRay and such are all based on those new 
and physically plausible (if not physically accurate) description of 
materials and light. What is more annoying is that it is impossible to map 
legacy material properties to physically plausible material properties.

There you have it. My profound refactoring in a "nutshell" =)

Though I'm not sure how a cache-friendly
> shader/ray tracing pipeline would work (are there any papers  that
> address a cache friendly shader and ray tracing pipeline, as opposed
> to just a ray tracing pipeline?).

That is a vast subject and it is difficult to point to one paper or eve a 
small set of paper that covers this topic.  I would say that the modern 
rendering papers are all concerned with figuring tricks for improving the 
hardware utilisation. That include the memory cache. But if there is one 
publication that I think is truely illuminating in this regard IMO, it would 
have to be "Heuristics Ray Shooting Algorithms" PhD thesis by Vlastimil 
Havran.

Yves