The Third Annual ICFP Programming Contest
(Version 1.18)

If you are viewing this document using Netscape, you may need to configure your browser to see the symbols properly. See for details. This document is also available in pdf and postscript formats.

1   The problem

This year's ICFP programming challenge is to implement a ray tracer. The input to the ray tracer is a scene description written in a simple functional language, called GML. Execution of a GML program produces zero, or more, image files, which are in PPM format. A web page of sample images, along with the GML inputs that were used to produce them, is linked off of the contest home page. The feature set of GML is organized into three tiers. Submissions must implement the first tier of features and extra credit will be given to submissions that implement the second or third tiers. Submissions will be evaluated on three scales: correctness of the produced images, run-time performance, and the tier of implemented GML features.

GML has primitives for defining simple geometric objects (e.g., planes, spheres, and cubes) and lighting sources. The surface properties used to render the objects are specified as functions in GML itself. In addition to supporting scene description, GML also has a render operator that renders a scene to an image file. For each pixel in the output image, the render command must compute a color. Conceptually, this color is computed by tracing the path of the light backwards from the eye of the viewer, to where it bounced off an object, and ultimately back to the light sources.

This document is organized as follows. Section 2 describes the syntax and general semantics of the modeling language. It is followed by Section 3, which describes those aspects of the language that are specific to ray tracing. Section 4 specifies the submission requirements and Section 5 provides hints about algorithms and pointers to online resources to get you started. The Appendix gives a summary of the operators in the modeling language.

This document is a bit on the long side because we have tried to make it complete and self-contained. (In fact, the LATEX source for this document is longer than our sample implementation!)

2   The modeling language

The input to the ray tracer is a scene description (or model) written in a functional modeling language called GML. The language has a syntax and execution model that is similar to PostScript (and Forth), but GML is lexically scoped and does not have side effects.

2.1   Syntax

A GML program is written using a subset of the printable ASCII character set (including space), plus tab, return, linefeed and vertical tab characters. The space, tab, return, linefeed and vertical tab characters are called whitespace.

The characters %, [, ], {, } are special characters.

Any occurrence of the character ``%'' not inside a string literal (see below) starts a comment, which runs to the end of the current line. Comments are treated as whitespace when tokenizing the input file.

The syntax of GML is given in Figure 1 (an opt superscript means an optional item and a * superscript means a sequence of zero or more items).

::= TokenGroup*
::= Token
| { TokenList }
| [ TokenList ]
::= Operator
| Identifier
| Binder
| Boolean
| Number
| String

Figure 1: GML grammar

A GML program is a token list, which is a sequence of zero or more token groups. A token group is either a single token, a function (a token list enclosed in `{' `}'), or an array (a token list enclosed in `[' `]'). Tokens do not have to be separated by white space when it is unambiguous. Whitespace is not allowed in numbers, identifiers, or binders.

Identifiers must start with an letter and can contain letters, digits, dashes (`-'), and underscores (`_'). A subset of the identifiers are used as predefined operators, which may not be rebound. A list of the operators can be found in the appendix. A binder is an identifier prefixed with a `/' character.

Booleans are either the literal true or the literal false. Like operators, true and false may not be rebound.

Numbers are either integers or reals. The syntax of numbers is given by the following grammar:
::= Integer
| Real
::= -opt DecimalNumber
::= -opt DecimalNumber . DecimalNumber Exponentopt
| -opt DecimalNumber Exponent
::= e -opt DecimalNumber
| E -opt DecimalNumber
where a DecimalNumber is a sequence of one or more decimal digits. Integers should have at least 24-bits of precision and reals should be represented by double-precision IEEE floating-point values.

Strings are written enclosed in double quotes (`"') and may contain any printable character other than the double quote (but including the space character). There are no escape sequences.

2.2   Evaluation

We define the evaluation semantics of a GML program using an abstract machine. The state of the machine is a triple <ENVac>, where ENV is an environment mapping identifiers to values, a is a stack of values, and c is a sequence of token groups. More formally, we use the following semantic definitions:
i in Int
i in BaseValue = Boolean È Int È Real È String
v in Value = BaseValue È Closure È Array È Point È Object È Light
(ENV, c) in Closure = Env # Code
a,[v1 ... vn] in Array = Value*
ENV in Env = Id --> Value
a,b in Stack = Value*
c in Code = TokenList
Evaluation from one state to another is written as <ENVac> ==> <ENV'; a'; c'> . We define ==>* to be the transitive closure of ==>. Figure 2 gives the GML evaluation rules.

<ENVai c> ==> <ENVa ic>     (1)
<ENVa v/x c> ==> <ENV±{x := v}; ac>     (2)
<ENVax c> ==> <ENVa ENV(x); c>     (3)
<ENVa{c'} c> ==> <ENVa (ENV, c'); c>     (4)
  <ENV'; ac'> ==>* <ENV''; b; Ø>  
  <ENVa (ENV', c'); apply c> ==> <ENVbc>  
  <ENV; Ø; c'> ==>* <ENV'; v1 ... vn; Ø>  
  <ENVa; [c'] c> ==> <ENVa [v1 ... vn]; c>  
  <ENV1ac1> ==>* <ENV''; b; Ø>  
  <ENVa true (ENV1, c1) (ENV2, c2);  if c> ==> <ENVbc>  
  <ENV2ac2> ==>* <ENV''; b; Ø>  
  < ENVa false (ENV1, c1) (ENV2, c2);  if c> ==> <ENVbc>  
  a  OPERATOR  a'  
  <ENVb aOPERATOR c> ==> <ENVb a'; c>  

Figure 2: Evaluation rules for GML

In these rules, we write stacks with the top to the right (e.g.; a x is a stack with x as its top element) and token sequences are written with the first token on the left. We use Ø to signify the empty stack and the empty code sequence.

Rule 1 describes the evaluation of a literal token, which is pushed on the stack. The next two rules describe the semantics of variable binding and reference. Rules 4 and 5 describe function-closure creation and the apply operator. Rule 6 describes the evaluation of an array expression; note that body of the array expression is evaluated on an initially empty stack. The semantics of the if operator are given by Rules 7 and 8. The last evaluation rule (Rule 9) describes how an operator (other than one of the control operators) is evaluated. We write
to mean that the operator OPERATOR transforms the stack a to the stack a'. This notation is used below to specify the GML operators.

We write Eval(c, v1, ..., vn) = (v'1, ..., v'n) for when a program c yields (v'1, ..., v'n) when applied to the values v1, ..., vn; i.e., when <{}; v1 ··· vnc> ==>* <ENVv'1 ···,v'n; Ø> .

There is no direct support for recursion in GML, but one can program recursive functions by explicitly passing the function as an extra argument to itself (see Section 2.7 for an example).

2.3   Control operators

GML contains two control operators that can be used to implement control structures. These operators are formally defined in Figure 2, but we provide an informal description here.

The apply operator takes a function closure, (ENV, c), off the stack and evaluates c using the environment ENV and the current stack. When evaluation of c is complete (i.e., there are no more instructions left), the previous environment is restored and execution continues with the instruction after the apply. Argument and result passing is done via the stack. For example:
1 { /x x x } apply addi
will evaluate to 2. Note that functions bind their variables according to the environment where they are defined; not where they are applied. For example the following code evaluates to 3:
1 /x          % bind x to 1
{ x } /f      % the function f pushes the value of x
2 /x          % rebind x to 2
f apply x addi
The if operator takes two closures and a boolean off the stack and evaluates the first closure if the boolean is true, and the second if the boolean is false. For example,
b { 1 } { 2 } if
will result in 1 on the top of the stack if b is true, and 2 if it is false

2.4   Numbers

GML supports both integer and real numbers (which are represented by IEEE double-precision floating-point numbers). Many of the numeric operators have both integer and real versions, so we combine their descriptions in the following:
n1 n2  addi/addf  n3
computes the sum n3 of the numbers n1 and n2.
r1  acos  r2
computes the arc cosine r2 in degrees of r1. The result is undefined if r1 < -1 or 1 < r1.
r1  asin  r2
computes the arc sine r2 in degrees of r1. The result is undefined if r1 < -1 or 1 < r1.
r1  clampf  r2
computes r2 = {
0.0 r1 < 0.0
1.0 r1 > 1.0
r1 otherwise
r1  cos  r2
computes the cosine r2 of r1 in degrees.
n1 n2  divi/divf  n3
computes the quotient n3 of dividing the number n1 by n2. The divi operator rounds its result towards 0. For the divi operator, if n2 is zero, then the program halts. For divf, the effect of division by zero is undefined.
n1 n2  eqi/eqf  b
compares the numbers n1 and n2 and pushes true if n1 is equal to n2; otherwise false is pushed.
r  floor  i
converts the real r to the greatest integer i that is less than or equal to r.
r1  frac  r2
computes the fractional part r2 of the real number r1. The result r2 will always have the same sign as the argument r1.
n1 n2  lessi/lessf  b
compares the numbers n1 and n2 and pushes true if n1 is less than n2; otherwise false is pushed.
i1 i2  modi  i3
computes the remainder i3 of dividing i1 by i2. The following relation holds between divi and modi:
i2 (idivi i2) + (imod i2) = i1
n1 n2  muli/mulf  n3
computes the product n3 of the numbers n1 and n2.
n1  negi/negf  n2
computes the negation n2 of the number n1.
i  real  r
converts the integer i to its real representation r.
r1  sin  r2
computes the sine r2 of r1 in degrees.
r1  sqrt  r2
computes the square root r2 of r1. If r1 is negative, then the interpreter should halt.
n1 n2  subi/subf  n3
computes the difference n3 of subtracting the number n2 from n1.

2.5   Points

A point is comprised of three real numbers. Points are used to represent positions, vectors, and colors (in the latter case, the range of the components is restricted to [0.0, 1.0]). There are four operations on points:
p  getx  x
gets the first component x of the point p.
p  gety  y
gets the second component y of the point p.
p  getz  z
gets the third component z of the point p.
x y z  point  p
creates a point p from the reals x, y, and z.

2.6   Arrays

There are two operations on arrays:
arr i  get  vi
gets the ith element of the array arr. Array indexing is zero based in GML. If i is out of bounds, the GML interpreter should terminate.
arr  length  n
gets the number of elements in the array arr.
The elements of an array do not have to have the same type and arrays can be used to construct data structures. For example, we can implement lists using two-element arrays for cons cells and the zero-length array for nil.
[] /nil
{ /cdr /car [ car cdr ] } /cons
We can also write a function that ``pattern matches'' on the head of a list.
{ /if-cons /if-nil /lst
  lst length 0 eqi
  { lst 0 get lst 1 get if-cons apply }

2.7   Examples

Some simple function definitions written in GML:
{ } /id                            % the identity function
{ 1 addi } /inc                    % the increment function
{ /x /y x y } /swap                % swap the top two stack locations
{ /x x x } /dup                    % duplicate the top of the stack
{ dup apply muli } /sq             % the squaring function
{ /a /b a { true } { b } if } /or  % logical-or function
{ /p                               % negate a point value
  p getx negf
  p gety negf
  p getz negf point
} /negp
A more substantial example is the GML version of the recursive factorial function:
{ /self /n
  n 2 lessi
  { 1 }
  { n 1 subi self self apply n muli }
} /fact
Notice that this function follows the convention of passing itself as the top-most argument on the stack. We can compute the factorial of 12 with the expression
12 fact fact apply

3   Ray tracing

In this section, we describe how the GML interpreter supports ray tracing.

3.1   Coordinate systems

GML models are defined in terms of two coordinate systems: world coordinates and object coordinates. World coordinates are used to specify the position of lights while object coordinates are used to specify primitive objects. There are six transformation operators (described in Section 3.3) that are used to map object space to world space.

The world-coordinate system is left-handed. The X-axis goes to the right, the Y-axis goes up, and the Z-axis goes away from the viewer.

3.2   Geometric primitives

There are five operations in GML for constructing primitive solids: sphere, cube, cylinder, cone, and plane. Each of these operations takes a single function as an argument, which defines the primitive's surface properties (see Section 3.6).
surface  sphere  obj
creates a sphere of radius 1 centered at the origin with surface properties specified by the function surface. Formally, the sphere is defined by x2 + y2 + z2 £ 1.
surface  cube  obj
creates a unit cube with opposite corners (0,0,0) and (1,1,1). The function surface specifies the cube's surface properties. Formally, the cube is defined by 0 £ x £ 1, 0 £ y £ 1, and 0 £ z £ 1. Cubes are a Tier-2 feature.
surface  cylinder  obj
creates a cylinder of radius 1 and height 1 with surface properties specified by the function surface. The base of the cylinder is centered at (0, 0, 0) and the top is centered at (0, 1, 0) (i.e., the axis of the cylinder is the Y-axis). Formally, the cylinder is defined by x2 + z2 £ 1 and 0 £ y £ 1. Cylinders are a Tier-2 feature.
surface  cone  obj
creates a cone with base radius 1 and height 1 with surface properties specified by the function surface. The apex of the cone is at (0, 0, 0) and the base of the cone is centered at (0, 1, 0). Formally, the cone is defined by x2 + z2 - y2 £ 0 and 0 £ y £ 1. Cones are a Tier-2 feature.
surface  plane  obj
creates a plane object with the equation y = 0 with surface properties specified by the function surface. Formally, the plane is the half-space y £ 0.

3.3   Transformations

Fixed size objects at the origin are not very interesting, so GML provides transformation operations to place objects in world space. Each transformation operator takes an object and one or more reals as arguments and returns the transformed object. The operations are:
obj rtx rty rtz  translate  obj'
translates obj by the vector (rtx, rty, rtz). I.e., if obj is at position (px, py, pz), then obj' is at position (px+rtx, py+rty, pz+rtz).
obj rsx rsy rsz  scale  obj'
scales obj by rsx in the X-dimension, rsy in the Y-dimension, and rsz in the Z dimension.
obj rs  uscale  obj'
uniformly scales obj by rs in each dimension. This operation is called Isotropic scaling.
obj q  rotatex  obj'
rotates obj around the X-axis by q degrees. Rotation is measured counterclockwise when looking along the X-axis from the origin towards +¥.
obj q  rotatey  obj'
rotates obj around the Y-axis by q degrees. Rotation is measured counterclockwise when looking along the Y-axis from the origin towards +¥.
obj q  rotatez  obj'
rotates obj around the Z-axis by q degrees. Rotation is measured counterclockwise when looking along the Z-axis from the origin towards +¥.
For example, if we want to put a sphere of radius 2.0 at (5.0, 5.0, 5.0), we can use the following GML code:
{ ... } sphere
2.0 uscale
5.0 5.0 5.0 translate
The first line creates the sphere (as described in Section 3.2, the sphere operator takes a single function argument). The second line uniformly scales the sphere by a factor of 2.0, and the third line translates the sphere to (5.0, 5.0, 5.0).

These transformations are all affine transformations and they have the property of preserving the straightness of lines and parallelism between lines, but they can alter the distance between points and the angle between lines. Using homogeneous coordinates, these transformations can be expressed as multiplication by a 4#4 matrix. Figure 3 describes the matrices that correspond to each of the transformation operators.

1 0 0
0 1 0
0 0 1
0 0 0 1
0 0 0
0 0
0 0
0 0 0 1
rs 0 0 0
0 rs 0 0
0 0 rs 0
0 0 0 1
Translation Scale matrix Isotropic scale matrix
1 0 0 0
0 cos(q) -sin(q) 0
0 sin(q) cos(q) 0
0 0 0 1
cos(q) 0 sin(q) 0
0 1 0 0
-sin(q) 0 cos(q) 0
0 0 0 1
cos(q) -sin(q) 0 0
sin(q) cos(q) 0 0
0 0 1 0
0 0 0 1
Rotation (X-axis) Rotation (Y-axis) Rotation (Z-axis)

Figure 3: Transformation matrices

For example, translating the point (2.6, 3.0, -5.0) by (-1.6, -2.0, 6.0) is expressed as the following multiplication:

1.0 0.0 0.0 -1.6
0.0 1.0 0.0 -2.0
0.0 0.0 1.0 6.0
0.0 0.0 0.0 1.0




Observe that points have a fourth coordinate of 1, whereas vectors have a fourth coordinate of 0. Thus, translation has no effect on vectors.

3.4   Illumination model

When the ray that shoots from the eye position through a pixel hits a surface, we need to apply the illumination equation to determine what color the pixel should have. Figure 4 shows a situation where a ray from the viewer has hit a surface.

Figure 4: A ray intersecting a surface

The illumination at this point is given by the following equation:
I = kd Ia C + kd
(N·L j) Ij C + ks
(N·Hj)n Ij C + ks Is C     (10)
C = surface color
Ia = intensity of ambient lighting
kd = diffuse reflection coefficient
N = unit surface normal
Lj = unit vector in direction of jth light source
Ij = intensity of jth light source
ks = specular reflection coefficient
Hj = unit vector in the direction halfway between the viewer and Lj
n = Phong exponent
Is = intensity of light from direction S
The view vector, N, and S all lie in the same plane. The vector S is called the reflection vector and forms same angle with N as the vector to the viewer does (this angle is labeled q in Figure 4). Light intensity is represented as point in GML and multiplication of points is component wise. The values of C, kd, ks, and n are the surface properties of the object at the point of reflection. Section 3.6 describes the mechanism for specifying these values for an object.

Computing the contribution of lights (the Ij part of the above equation) requires casting a shadow ray from the intersection point to the light's position. If the ray hits an object that is closer than the light, then the light does not contribute to the illumination of the intersection point.

Ray tracing is a recursive process. Computing the value of Is requires shooting a ray in the direction of S and seeing what object (if any) it intersects. To avoid infinite recursion, we limit the tracing to some depth. The depth limit is given as an argument to the render operator (see Section 3.8).

3.5   Lights

GML supports three types of light sources: directional lights, point lights and spotlights. Directional lights are assumed to be infinitely far away and have only a direction. Point lights have a position and an intensity (specified as a color triple), and they emit light uniformly in all directions. Spotlights emit a cone of light in a given direction. The light cone is specified by three parameters: the light's direction, the light's cutoff angle, and an attenuation exponent (see Figure 5).

Figure 5: Spotlight

Unlike geometric objects, lights are defined in terms of world coordinates.

dir color  light  l
creates a directional light source at infinity with direction dir and intensity color. Both dir and color are specified as point values.
pos color  pointlight  l
creates a point-light source at the world coordinate position pos with intensity color. Both pos and color are specified as point values. Pointlights are a Tier-2 feature.
pos at color cutoff exp  spotlight  l
creates a spotlight source at the world coordinate position pos pointing towards the position at. The light's color is given by color. The spotlight's cutoff angle is given in degrees by cutoff and the attenuation exponent is given by exp (these are real numbers). The intensity of the light from a spotlight at a point Q is determined by the angle between the light's direction vector (i.e., the vector from pos to at) and the vector from pos to Q. If the angle is greater than the cutoff angle, then intensity is zero; otherwise the intensity is given by the equation
I =



color     (11)
Spotlights are a Tier-3 feature.
The light from point lights and spotlights is attenuated by the distance from the light to the surface. The attenuation equation is:
100 I
99 + d2
where d is the distance from the light to the surface and I is the intensity of the light. Thus at a distance of 5 units the strength of the light will be about 85% and at 10 units it will be about 50%. Note that the light reflected from surfaces (the ks Is C term in Equation 3.4) is not attenuated; nor is the light from directional sources.

3.6   Surface functions

GML uses procedural texturing to describe the surface properties of objects. The basic idea is that the model provides a function for each object, which maps positions on the object to the surface properties that determine how the object is illuminated (see Section 3.4).

A surface function takes three arguments: an integer specifying an object's face and two texture coordinates. For all objects, except planes, the texture coordinates are restricted to the range 0 £ u,v £ 1. The Table 1 specifies how these coordinates map to points in object-space for the various builtin graphical objects.

Table 1: Texture coordinates for primitives

(0, u, v) (sqrt(1 - y2)sin(360 u), y, sqrt(1 - y2)cos(360 u)), where y = 2 v - 1
(0, u, v) (u, v, 0) front
(1, u, v) (u, v, 1) back
(2, u, v) (0, v, u) left
(3, u, v) (1, v, u) right
(4, u, v) (u, 1, v) top
(5, u, v) (u, 0, v) bottom
(0, u, v) (sin(360 u), v, cos(360 u)) side
(1, u, v) (2 u - 1, 1, 2 v - 1) top
(2, u, v) (2 u - 1, 0, 2 v - 1) bottom
(0, u, v) (v sin(360 u), v, v cos(360 u)) side
(1, u, v) (2 u - 1, 1, 2 v - 1) base
(0, u, v) (u, 0, v)

Note that (as always in GML), the arguments to the sin and cos functions are in degrees. The GML implementation is responsible for the inverse mapping; i.e., given a point on a solid, compute the texture coordinates.

A surface function returns a point representing the surface color (C), and three real numbers: the diffuse reflection coefficient (kd), the specular reflection coefficient (ks), and the Phong exponent (n). For example, the code in Figure 6 defines a cube with a matte 3#3 black and white checked pattern on each face.

0.0 0.0 0.0 point /black
1.0 1.0 1.0 point /white

[                                 % 3x3 pattern
  [ black white black ]
  [ white black white ]
  [ black white black ]
] /texture

{ /v /u /face                     % bind parameters
  {                               % toIntCoord : float -> int
    3.0 mulf floor /i               % i = floor(3.0*r)
    i 3 eqi { 2 } { i } if           % make sure i is not 3
  } /toIntCoord
  texture u toIntCoord apply get  % color = texture[u][v]
    v toIntCoord apply get
  1.0                             % kd = 1.0
  0.0                             % ks = 0.0
  1.0                             % n = 1.0
} cube
Figure 6: A checked pattern on a cube

3.7   Constructive solid geometry

Solid objects may be combined using boolean set operations to form more complex solids. There are three composition operations:
obj1 obj2  union  obj3
forms the union obj3 of the two solids obj1 and obj2.
obj1 obj2  intersect  obj3
forms the intersection obj3 of the two solids obj1 and obj2. The intersect operator is a Tier-3 feature.
obj1 obj2  difference  obj3
forms the solid obj3 that is the solid obj1 minus the solid obj2. The difference operator is a Tier-3 feature.
We can determine the intersection of a ray and a compound solid by recursively computing the intersections of the ray and the solid's pieces (both entries and exits) and then merging the information according to the boolean composition operator. Figure 7 illustrates this process for two objects (this picture is called a Roth diagram).

Figure 7: Tracing a ray through a compound solid

When rendering a composite object, the surface properties are determined by the primitive that defines the surface. If the surfaces of two primitives coincide, then which primitive defines the surface properties is unspecified.

3.8   Rendering

The render operator causes the scene to be rendered to a file.
amb lights obj depth fov  wid ht file   render  ---
The render operator renders a scene to a file. It takes eight arguments:
the intensity of ambient light (a point).
is an array of lights used to illuminate the scene.
is the scene to render.
is an integer limit on the recursive depth of the ray tracing owing to specular reflection. I.e., when depth = 0, we do not recursively compute the contribution from the direction of reflection (S in Figure 4).
is the horizontal field of view in degrees (a real number).
is the width of the rendered image in pixels (an integer).
is the height of the rendered image in pixels (an integer).
is a string specifying output file for the rendered image.
The render operator is the only GML operator with side effects (i.e., it modifies the host file system). A GML program may contain multiple render operators (for animation effects), but the order in which the output files are generated is implementation dependent. The results of evaluating the render operator during the evaluation of a surface function are undefined (i.e., your program may choose to exit with an error, or execute the operation, or do something else).

When rendering a scene, the eye position is fixed at (0, 0, -1) looking down the Z-axis and the image plane is the XY-plane (see Figure 8). The horizontal field of view (fov) determines the width of the image in world space (i.e., it is 2 tan(0.5 fov)), and the height is determined from the aspect ratio. If the upper-left corner of the image is at (x, y, 0) and the width of a pixel is D, then the ray through the jth pixel in the ith row has a direction of (x + (j+0.5)D, y - (i+0.5)D, 1).

Figure 8: View coordinate system

When the render operation detects that a ray has intersected the surface of an object, it must compute the texture coordinates at the point of intersection and apply the surface function to them. Let (face, u, v) be the texture coordinates and surf be the surface function at the point of intersection, and let
Eval(surf apply, face, u, v) = (C, kd, ks, n)
Then the surface properties for the illumination equation (see Section 3.4) are C, kd, ks, and n.

3.9   The output format

The output format is the Portable Pixmap (PPM) file format.1 The format consists of a ASCII header followed by the pixel data in binary form. The format of the header is These items are separated by whitespace (blanks, TABs, CRs, and LFs). After the maximum color value, there is a single whitespace character (usually a newline), which is followed by the pixel data. The pixel data is a sequence of three-byte pixel values (red, green, blue) in row-major order. Light intensity values (represented as GML points) are converted to RGB format by clamping the range and scaling.

In the header, characters from a ``#'' to the next end-of-line are ignored (comments). This comment mechanism should be used to include the group's name immediately following the line with the magic number. For example, the sample implementation produces the following header:
# GML Sample Implementation
256 256

4   Requirements

Your program should take its input from standard input (i.e., UNIX file descriptor 0). Execution of the input specification will result in zero or more images being rendered to files. If your implementation detects an error, it should return a non-zero exit status; otherwise it should return a zero exit status upon successful termination. Our test harness relies on this error status being set correctly, so be sure to get them right!

Your program should detect syntactically incorrect input and run-time type errors (the latter may be detected statically, if you wish). It should also catch array accesses that are out of range. Other errors, such as integer overflows and division by zero, may be detected and reported, but it is not necessary. In particular, implementations are free to generate NaNs and Infs when doing floating-point computations.

The submission requirements are described in detail at, but we summarize them here. Your submission should include a README file containing a brief description of the submission, programming language(s) used, and anything else that you want to bring to the attention of the judges.

Submissions will be evaluated on their correctness, speed of execution, and set of implemented GML features. For the latter metric, we have grouped the features of GML into three tiers as follows:
Tier 1
The first tier consists of the operations described in Section 2, plus planes, spheres, and directional lights. All GML operators except cone, cube, cylinder, difference, intersect, pointlight, and spotlight should be implemented.
Tier 2
This tier adds more primitive solids and additional lighting to Tier 1. The additional operators are: cone, cube, cylinder, and pointlight.
Tier 3
This tier adds constructive solid geometry and additional lighting to Tier 2. The additional operators are: difference, intersect, and spotlight.
Your README file should specify which tier your submission implements.

Judging of the contest entries will proceed in three phases. First, we will evaluate each submission for basic correctness using very simple Tier-1 test cases. Programs that fail to run, dump core, etc. will be disqualified at the end of this phase. The second phase tests the basic correctness of submissions (without regards to performance). We will use a selection of Tier-1 test cases and compare the output with that generated by our sample implementations. Submissions that deviate significantly from the the reference outputs will be disqualified. The third phase will compare the performance and implemented features of the submissions. When comparing submissions, a program that implements Tier-1 will have to be significantly faster than a Tier-2 program to beat it. Likewise, a Tier-2 program will have to be significantly faster than a Tier-3 program to beat it. Image quality also matters; for example, a program that has surface acne will be penalized. Consideration will be given for interesting sample images.

5   Hints

5.1   Basic facts

The dot product of two vectors v1 = (x1, y1, z1) and v2 = (x2, y2, z2) is v1·v2 = (x1 x2 + y1 y2 + z1 z2). When v1 and v2 are unit vectors, then v1·v2 is the cosine of the angle formed by the two vectors. More generally, v1·v2 = |v1| |v2| cos(q), where q is the angle between the vectors.

5.2   Intersection testing

A plane P can be defined by its unit normal Pn and the distance d from the plane to the origin. The half-space that P = (Pn, d) defines are those points Q such that Q·Pn + d £ 0. Given this definition, the intersection of a ray R(t) = (Ro + t Rd) and a plane (Pn, d) is given by the equation
-(Pn · Ro + d)
If Pn·Rd = 0, then the ray is parallel to the plane (it might lie in the plane, but we can ignore that case for our purposes). If tintersection < 0, then the line defined by the ray intersects the plane behind the ray's origin; otherwise the point of intersection is R(tintersection). We can tell which side of the plane Ro lies by examining the sign of Pn·Rd; if it is positive, then Ro is in the half-space defined by P.

Computing the intersection of a ray R(t) = (Ro + t Rd) and a sphere S centered at Sc with radius r is more complicated. Let loc be the length of the vector from the ray's origin to the center of the sphere; then if loc < r, the ray originates inside the sphere. We can compute the distance along the ray from the ray's origin to the closest approach to the sphere's center by the equation tca = (Sc - Ro)·Rd (see Figure 9). If tca < 0, then the ray is pointing away from the sphere's center, which means that if the ray's origin is outside the sphere then there is no intersection. Once we have computed tca, we can compute the square of the distance from the ray to the center at the point of closest approach by the d2 = loc2 - tca2. From this, we can compute the square of the half chord distance thc2 = r2 - d2 = r2 - loc2 + tca2. As can be seen in Figure 9, if thc<0, then the ray does not intersect the sphere, otherwise the points of intersection are given by R(tca±thc) (assuming the ray originates outside the sphere).

Figure 9: Ray/sphere intersection

The intersection of a ray and a cube can be determined by using the technique given for planes (test against the planes containing the faces of the cube). Intersections for cones and cylinders can be determined by plugging the ray equation (R(t) = Ro + t Rd) into the equations for the surface. In both cases (as for spheres) the solution requires pluggin values into the quadratic formula.

One approach to ray tracing with a modeling language that supports affine transformations (such as GML) is to transform the rays into object space and do the intersection tests there. This approach allows the intersection tests to be specialized to the standard objects, which can greatly simplify the tests. Remember, however, that affine transformations do not preserve lengths --- applying an affine transformation to a unit vector will not yield a unit vector in general.

5.3   Surface acne

One problem that you are likely to encounter is called surface acne and results from precision errors. The problem arises from when the origin of a shadow ray is on the wrong side of its originating surface, and thus intersets the surface. The visual result is usually a black dot at that pixel. The sample images include an example that illustrates this problem. One solution is to offset the shadow ray's origin by a small amount in the ray's direction. Another solution is not to test intersection's against the originating surface.

5.4   Optimizations

There are opportunities for performance improvements both in the the implementation of the GML interpreter and in the ray tracing engine.

While the time spent to compute the objects in a scene is typically small compared to the rendering time, the GML functions that define the surface properties get evaluated for every ray intersection. You may find it useful to analyse surface functions for the common case where they are constant.

The resources listed below include information on techniques for improving the efficiency of ray tracing. Most of these techniques focus on reducing the cost or number of ray/solid intersection tests. For example, if you precompute a bounding volume for a complex object, then a quick test against the bounding volume may allow you to avoid a more expensive test against the object. If your implementation supports the Tier-3 CSG operators, then you probably want to have a version of your intersection testing code that is specialized for shadow rays.

5.5   Resources

Here are a few pointers to on-line sources of information about graphical algorithms and ray tracing.

is the ICFP'00 contest home page.

is a page of example GML specifications with the expected images.

is a text file that lists all of the GML operators.

is the 3D Object Intersection page with pointers to papers and code describing various intersection algorithms.

is the home page of the Ray Tracing News, which is an online journal about ray tracing techniques.

is a paper by Brian Smits on efficiency issues in implementing ray tracers.

is the source-code repository for the Graphics Gems series.

is the FAQ for the news group.

has source code for various graphical algorithms.

Operator summary

The following is an alphabetical listing of the GML operators with brief descriptions. The third column lists the section where the operator is defined and the fourth column specifies which implementation tier the operator belongs to.
Name Description Section Tier
acos arc cosine function 2.4 *
addi integer addition 2.4 *
addf real addition 2.4 *
apply function application operator 2.3 *
asin arc sine function 2.4 *
clampf clamp the range of a real number 2.4 *
cone a unit cone 3.2 **
cos cosine function 2.4 *
cube a unit cube 3.2 **
cylinder a unit cylinder 3.2 **
difference difference of two solids 3.7 ***
divi integer division 2.4 *
divf real division 2.4 *
eqi integer equality comparison 2.4 *
eqf real equality comparison 2.4 *
floor real to integer conversion 2.4 *
frac fractional part of real number 2.4 *
get get an array element 2.6 *
getx get x component of point 2.5 *
gety get y component of point 2.5 *
getz get z component of point 2.5 *
if conditional control operator 2.3 *
intersect intersection of two solids 3.7 ***
length array length 2.6 *
lessi integer less-than comparison 2.4 *
lessf real less-than comparison 2.4 *
light defines a directional light source 3.5 *
modi integer remainder 2.4 *
muli integer multiplication 2.4 *
mulf real multiplication 2.4 *
negi integer negation 2.4 *
negf real negation 2.4 *
plane the XZ-plane 3.2 *
point create a point value 2.5 *
pointlight defines a point-light source 3.5 **
real convert an integer to a real number 2.4 *
render render a scene to a file 3.8 *
rotatex rotation around the X-axis 3.3 *
rotatey rotation around the Y-axis 3.3 *
rotatez rotation around the Z-axis 3.3 *
scale scaling transform 3.3 *
sin sine function 2.4 *
sphere a unit sphere 3.2 *
spotlight defines a spotlight source 3.5 ***
sqrt square root 2.4 *
subi integer subtraction 2.4 *
subf real subtraction 2.4 *
translate translation transform 3.3 *
union union of two solids 3.7 *
uscale uniform scaling transform 3.3 *

Change history

A bunch of HTML rendering workarounds.
Description of how surface functions are applied was missing the face argument.
Corrected sloppy language about illumination vectors.
Clarified who rendering depth limit works; corrected text about light attenuation; and fixed texture equations for cone and cylinder end caps.
Got the attenuation equation fix into the document this time.
Clarified definition of modi; fixed typo in description of initial ray direction; clarified types of light operators; corrected typo in attenuation equation (should be d2, not d3); and added note about conversion to RGB format.
Added note about number sizes and fixed texture coordinates of planes.
Many fixes: added specification of the render operation's types; fixed typo in definition of dot product; added clarification about illumination equation and vector multiplication; fixed typo in equation for square of half-chord distance; and fixed texture coordinate equations for spheres and cones.
Clarified definition of frac operator.
Added note about rebinding true and false.
Added discussion about applying render in a surface function.
Fixed inc example.
Fixed swap example.
Fixed typo in divi/divf description; added text to clarify syntax.
Fixed mistake in factorial example.
Added version number and change history.
Fixed rule cross references in HTML version.
Fixed bug in example; sub should have been get.
First release.

The xv program, available on most Unix systems, and the IrfanView viewer for Microsoft Windows (available from both understand the PPM format.

This document was translated from LATEX by HEVEA.