POV-Ray

There is an issue where the pause button in POVWIN will sometimes not work during an animation (primarily where the frame rate is high), and furthermore, POVWIN can then get into a state where it’s not possible to use the pause until it is re-started.

Newsgroup report.

Windows UI code should be refactored to use _TCHAR throughout instead of char, as well as the corresponding string function macros, to head for Unicode support.

Subsurface scattering must be made photon-aware.

Near concave edges, radiosity samples may be re-used at a longer distance away from the edge than towards the edge; there is code in place to ensure this, but it only works properly where two surfaces meet roughly rectangularly, while failing near the junction of three surfaces or non-rectangular edges, potentially causing “cross-talk”.

It should be investigated how the algorithm can be improved or replaced to better cope with non-trivial geometry.

Some features are optional in custom builds of POV-Ray (I’m thinking about OpenEXR in particular); it would be nice to have a syntax for an SDL script to check for support of such features, so it may take some fallback action if the feature is not supported.

Background

Possible solution

The current 32-bit linear congruential generator used as RNG in POV-Ray is sometimes quite limited for some purposes and in a few cases its poor quality shows up (as has been demonstrated more than once in the newsgroup). Thus it would be nice if POV-Ray offered additional, higher-quality random number generators, besides the current one (which should probably remain for backwards compatibility). These RNGs could include algorithms like the Mersenne Twister and the ISAAC RNG, both of which have very decent quality and have an enormous periods (while at the same time being very fast).

After a long discussion, the following syntax for specifying the RNG type and seed (which may be larger than 32 bits) has been suggested:

seed(<value>) | seed(<type>, <value> [, <values>])

For example:

#declare Seed1 = seed(123); // Use the current RNG, with seed 123
#declare Seed2 = seed(1, 123); // Identical to the previous one
#declare Seed3 = seed(2, 456, 789, 123); // Use RNG algorithm #2,
                                         // with a large seed (96 bits specified here)

A C++ implementation of the ISAAC RNG can be found at http://warp.povusers.org/IsaacRand.zip

When you instantiate an object several times, eg:

object { MyObj translate -x*10 }
object { MyObj translate x*10 }

POV-Ray will copy that object in memory, at least for most types of objects. Not for all of them, though. Most famously if MyObj is a mesh, it won’t be copied, but only a reference to the original will be used, thus saving memory. (There are a few other primitives which also don’t cause a copy, such as bicubic_patch and blob, but those are naturally not so popular as mesh, so it’s a less known fact.)

AFAIK the reason why referencing (rather than copying) is not used for all types of objects is rather complicated, and mostly related to how transformations are applied to these objects. For example if the object being instantiated is a union, the translates above will be (AFAIK) applied to the individual members of the union rather than to the union object itself.

Copying, however, can be quite detrimental in some situations. For example if you have a huge union, and you want to instantiate it many times, the memory usage will be that many times larger (compared to just one instance). This is sometimes something which the user would not want, even if it made the rendering slightly slower as a consequence. (In other words, better to be able to render the scene in the first place, rather than running out of memory.)

Redesigning POV-Ray so that all objects would be referenced rather than copied would probably be a huge job, and in some cases a questionable one. There probably are situations where the current method really produces faster rendering times, so redesigning POV-Ray so that it would always reference instead of copy, could make some scenes render slower.

So this got me thinking about an alternative approach: How hard would it be to create a special object which sole purpose is to act as a reference to another object, without copying it? This special reference object would act as any regular object, would have its own transformation matrix and all that data related to objects, but its sole purpose is to simply be a “wrapper” which references an existing object. It could be, for example, like this:

object_ref { MyObj translate -x*10 }
object_ref { MyObj translate x*10 }

The end result would be exactly identical as earlier, but the difference is that now MyObj behaves in the same way as a mesh (in the sense that it’s not instantiated twice, but only once, even though it appears twice in the scene), regardless of what MyObj is.

In some cases this might render slightly slower than the first version (because POV-Ray has to apply the transformations of the object_ref first, after which it applies whatever transformations are inside MyObj), but that’s not the point here. The point is to save memory if MyObj is large.

An object_ref would behave like any other object, so you could do things like:

#declare MyObjRef = object_ref { MyObj };

object { MyObjRef translate -x*10 }
object { MyObjRef translate x*10 }

(The only thing being instantiated (and copied) here is the “MyObjRef” object, not the object it’s referring to, so that actual object is still stored in memory only once.)

In some situations it might even be so that referenced objects actually render faster than if the objects were copied because references increase data locality, lessening cache misses.

I believe this could be a rather useful feature and should be seriously considered, unless there are some major obstacles in implementing it.

There is an big issue with the slope pattern: when the object it is applied to is instanced (again) with a transformation (in particular a rotation, as a translation would not impact.. but a shear might), the colours of the surfaces are changed.

  
object { p translate -5*x }
object { p rotate 220*y+20*x    translate 3*x }       

Nobody would expect the object to be different in appearance.
If slope {} is replaced with wood, all is fine. (as for others textures, i guess)

IMHO, the slope vector need to be adjusted for the later transformation(s) (so as to compensate the issue of using the Perturbed Normal vector).

This should not impact the AOI/FACING (experimental) patterns, as AOI definition is pretty clear about duplicating & transform if you think about it a bit, as well as FACING: for these two, it is expected to either use the ray(current point of view) or a fixed 3D point as reference. At the limit, discussion about moving the 3D point of FACING might also be opened to interpretation.

AOI/FACING are in task #19

Front- & Back-end code should be refactored for full Unicode support in scene files and strings.

Currently, array items may only be referenced by their index number (an integer). It would be nice to also be able to assign string values as array indexes, as in other scripting languages.

Someone built a version of Povray with internal support for automatic subdivision of meshes. See:

http://www.cise.ufl.edu/~xwu/Pov-Sub/

Would like to see this feature added natively to Povray.

When rendering in multiple passes (radiosity in my case), the elapsed pixels and percentage, written to terminal
are first displayed like this:
Rendered 126202 of 360000 pixels (35%)
Then on the second stage the output text becomes shorter and you see
Rendered 25344 of 360000 pixels (7%)%)
The contents of the previous status are not erased, so the longer text persists (note the duplicate percentage sign and closing parenthesis). Such a glitch could have more drastic effect in rare cases.

I’m running
Version 3.7.0.RC3 (g++ 4.6.2 x86_64-unknown-linux-gnu)
compiled for the Arch Linux package.

The so-called shadow line artifact (http://wiki.povray.org/content/Knowledgebase:The_Shadow_Line_Artifact) which affects objects with a ‘normal’ statement as well as smooth meshes and heightfields can be really annoying sometimes. Currently the only way to remove it is to make the object shadowless, which isn’t a good solution except in very special cases.

This algorithm could remove the artifact: If the actual normal vector of the object points away from the light source (its dot-product with the light vector is negative) but the perturbed normal points towards it (dot-product positive), then ignore the first shadow-test intersection with the object itself.

There are alternative ways of implementing an equivalent functionality:

- Don’t check the condition (if it’s too difficult to check due to how the code is designed) but always ignore the first intersection with the objects itself. This will work properly with closed surfaces but not with open ones, so it might need to be a feature for the user to turn on with a keyword (similar to eg. ‘double_illuminate’).

- Alternatively, don’t ignore the first intersection, but instead ignore the “opposite side” of the object’s surface (again, possibly only if a keyword has been specified). In other words, if we are rendering the outer side of the object, ignore its inner side when shadow-testing, and vice-versa.

- Perhaps simply add a feature to make surfaces one-sided (similarly to how they can be made so in OpenGL and similar scanline rendering systems). In other words, the inner side of a surface is completely ignored everywhere, making the object virtually invisible from the inside. The advantage of this feature would be that it can have uses other than simply removing the shadow line artifact.

The sphere_sweep renders well when specified directly, but when it is scaled, its bounding box is calculated incorrectly, which clips the object so it almost disappears.

The effect is present for all three types of splines.

I’m attaching a test scene and the rendering result. The saving of the object with #declare has no effect, I just wanted to show both transformed and untransformed version.

I don’t think this issue is related to other artifacts occuring with sphere_sweep, as it is obviously an issue of the internal bounding box.

With very fast renders and very large output files, the message queue can fill up because the producers are not limited by IO, while the consumer performance is limited by disk IO. Consequently, the message queue can fill up to exhaust all available memory. The solution is to build in some better control of pending output data in the message queue on the producer side. This will also pave the way for message communication over slow links (i.e. a network).

In the photons map, the direction of each photon is stored as separate latitude & longitude angles (encoded in one byte each), causing the longitudinal direction component’s precision to be unnecessarily high for directions close to the “poles” (Y axis); in addition, encoded value -128 is never used. For better overall precision as well as precision homogenity, the following scheme could be used instead:

• Encode the latitude (-pi/2 to +pi/2) into LatCount=226 distinct values (= 256*sqrt(pi)/2) rounded to the next even number) from 0 to LatCount-1 using

latCode = (int)((LatCount-1) * (lat/M_PI + 0.5) + 0.5)

• For each latitude code, define a specific number of encodable longitude values, LngCount[latCode] = approx. cos(lat)*pi*65536/(2*LatCount); this can be a pre-computed table, and may need slight tweaking for optimum use of the code space. Encode the longitude (-pi to +pi) into a value from 0 to (LngCount[lat]-1) using

LC = LngCount[latCode];
lngCode = (int)(LC * (lng/(2*M_PI) + 0.5) + 0.5) % LC;

• Besides LngCount[latCode], also store the sum of LngCount[i] with i < latCode as LatBase[latCode]; encode the direction as

dirCode = LatBase[latCode] + lngCode;

• For decoding, a simple lookup from a precomputed list of directions could be used (2^15 entries, i.e. one hemisphere, will suffice). To conserve space, direction vectors could be scaled by (2^N-1) and stored as (N+1)-bit signed integer triples rather than floating point values; due to the limited precision of the lat/long information, 8 bits per coordinate might be enough, giving a table size of 96k. A full double-precision table would require 786k instead.

While familiarizing myself with the code, I found some small changes in the solve_cubic function that lead to a significant speedup.

In my experience, “pow” is by far the slowest function in math.h and replacing it with simpler functions usually makes a tremendous impact on the speed (it’s an order of magnitude slower than sqrt/exp/cbrt/log).

solve_cubic has a “pow” function that can be replaced by cbrt (cubic root), which is standard in ISO-C99 and should be available on all systems. Separate benchmarks of solve_cubic function show this change almost doubles the speed and does not lower the accuracy. As solve_cubic is part of the solution of quartic equation, this improves the speed for many primitives. Testing with a scene containing many torus intersection tests (attached below) I still observed almost 10% speedup (Intel, 4 threads, 2 hyperthreaded cores, antialiasing on, 600×600: from 91 to 84 seconds). And this is for a torus, where a lot of time is spent in the solve_quartic and cubic solver is only called once! Similar speedup should be expected for prism, ovus, sor and blob.

I do believe the cubic solver can be done without trigonometry, but that would mean changing the algorithm, introducing new bugs and requiring a lot of testing. However, the trigonometric evaluation can still be simplified (3% speedup in full torus benchmark).

These changes don’t affect the algorithm at all, they are mathematically identical to the existing code, so the changes can be applied immediately. I also included other changes just as suggestions. Every change is commented and marked with [SC 2.2013].

This sadly does not speedup the sturm solver, which uses bisection and regula-falsi and looks very optimized already.

The test scene I used has a lot of torus intersections from various directions (shadow rays, main rays, transmitted rays).

In version 3.7.0.RC7.msvc10.win64, there is a bug in rendering Bézier patches in which four points (along one edge) are all the same point.

The rendering can be seen here:
http://i.imgur.com/eq2UIXR.png [Edit: See attachment for the rendering]

As you can see, there is a visible unwanted artifact in the corner of each patch. The two patches shown are essentially the same, except with the 4×4 matrix of vertices transposed (just to demonstrate that simply transposing it didn’t fix it).

Expected rendering is a smooth surface without the artifact.

Below is the code used to render the above example.

#version 3.7;

global_settings { assumed_gamma 1.0 }

camera {

  location <45, 31, -10>
  look_at <40, 21, 200>
  right x*image_width/image_height

}

light_source {

  <660, 300, -525>
  color rgb 1

}

Example 1: First point in each row is the same point
bicubic_patch {
type 1 flatness 0.001
u_steps 4 v_steps 4
<32.2168, -23.78125, 0>, <34.4968, -23.78125, 0>, <35.2168, -23.78125, -0.72>, <35.2168, -23.78125, -3>,
<32.2168, -23.78125, 0>, <34.4968, -22.10256, 0>, <35.2168, -21.57244, -0.72>, <35.2168, -21.57244, -3>,
<32.2168, -23.78125, 0>, <33.9709, -21.55577, 0>, <34.52483, -20.85299, -0.72>, <34.52483, -20.85299, -3>,
<32.2168, -23.78125, 0>, <32.30556, -21.50298, 0>, <32.33359, -20.78352, -0.72>, <32.33359, -20.78352, -3>
rotate 180*x scale 1.4
translate ←5, 0, 0>
pigment { color <1, 0, 0> }
} Example 2: First row is all the same point
bicubic_patch {

  type 1 flatness 0.001
  u_steps 4 v_steps 4
  <32.2168, -23.78125, 0>, <32.2168, -23.78125, 0>, <32.2168, -23.78125, 0>, <32.2168, -23.78125, 0>,
  <34.4968, -23.78125, 0>, <34.4968, -22.10256, 0>, <33.9709, -21.55577, 0>, <32.30556, -21.50298, 0>,
  <35.2168, -23.78125, -0.72>, <35.2168, -21.57244, -0.72>, <34.52483, -20.85299, -0.72>, <32.33359, -20.78352, -0.72>,
  <35.2168, -23.78125, -3>, <35.2168, -21.57244, -3>, <34.52483, -20.85299, -3>, <32.33359, -20.78352, -3>
  rotate 180*x

  scale 1.4
  pigment { color <1, 1, 0> }

}

When rendering only a region of a file, using the command-line options +sc/+sr/+ec/+er, the area of the image that is excluded comes out as black in the final PNG.

Expected behaviour is for it to be transparent.

It appears as a side effect of investigation of #294: the code in isosurf.cpp, inside
bool IsoSurface::Function_Find_Root_R(ISO_ThreadData& itd, const ISO_Pair* EP1, const ISO_Pair* EP2, DBL dt, DBL t21, DBL len, DBL& maxg)

	if(gradient < temp)
		gradient = temp;

is not thread-safe (The code is used at render time, there is a data race between < and = operation, as gradient is stored in the global object and accessed in write mode by the cited code)

It is only important if the gradient is initially undervaluated (otherwise, all is fine, no write-access)

As emerged as an idea during the discussion of FS#299, an SDL / POV-Ray editor feature would be useful that allows API documentation via formal comments, e.g. in include files:

/**
 * Creates a car object.
 * @param a
 *        description of param a
 * ...
 */
#macro car(a,b,c)
  ...
#end

In addition to the ability of (auto-)generating a documentation file from such comments, an editor window feature would be convenient that allows popup display of a macro’s (object’s / parameter’s / ...) documentation section.

When using OpenEXR output file format, POV-Ray erroneously reports it as “24 bpp EXR” in the message output, while in fact it generates a 3×16 = 48 bpp file.