STANDARD PROCEDURAL DATABASES, by Eric Haines, 3D/Eye, Inc. [Created while under contract to Hewlett-Packard FSD and HP Laboratories] Version 2.4, as of 5/1/88 address: 3D/Eye, Inc., 410 East Upland Road, Ithaca, NY 14850 e-mail: hpfcla!hpfcrs!eye!erich@hplabs.HP.COM History ------- This software package by Eric Haines is not copyrighted and can be used freely. Versions 1.0 to 1.5 released February to June, 1987 for testing. Version 2.0 released July, 1987. Version 2.1 released August, 1987 - corrected info on speed of the HP 320, other minor changes to README. Version 2.2 released November, 1987 - shortened file names to <=12 characters, procedure names to <= 32 characters, and ensured that all lines are <= 80 characters (including return character). Version 2.3 released March, 1988 - corrected gears.c to avoid interpenetration, corrected and added further instructions and global statistics for ray tracing to README. Version 2.4 released May, 1988 - fixed hashing function for mountain.c. {This file uses tab characters worth 8 spaces} Introduction ------------ This software is meant to act as a set of basic test images for ray tracing and hidden surface algorithms. The programs generate databases of objects which are fairly familiar and "standard" to the graphics community, such as a fractal mountain, a tree, a recursively built tetrahedral structure, etc. I created them originally for my own testing of a ray-tracer. My hope is that these will be used by researchers to test algorithms. In this way, research on algorithmic improvements can be compared on a more standard set of measures. At present, one researcher ray-traces a car, another a tree, and the question arises "How many cars to the tree?" With these databases we may be comparing oranges and apples ("how many hypercubes to a timeshared VAX?"), but it sure beats comparing oranges and orangutans. Another use for these databases is to find how many vectors or polygons per second a particular machine can actually generate. Talking with people at SIGGRAPH '86, I found that various vendors would say "well, they claim a rate of 5000 polygons/second, but their polygons are not our polygons". With these tests, their polygons are our polygons. I admit a bias towards the structure of the HP-SRX ("Renaissance") display system and its display software "Starbase" in the way I structured the testing conditions (laid out later), but this comes fairly close to preliminary PHIGS+ standards. For example, these databases are not kind to machines using the painter's algorithm, but I think this algorithm is inappropriate for good, general 3D modeling hardware. These databases do not use polygons where an intensity or color is given for each vertex, though such polygons are a feature of some machines (including the SRX). They do include polygonal patches, where a normal is given for each vertex, as this feature seems fairly common. Basically, if display hardware cannot provide 8 light sources, use of normals at vertices, specular shading, z-buffering, and perspective transformations, then I consider this machine somewhat ineffective for use in a 3D modeling environment (which is mostly all I care about--my personal bias). If you do not, please tell me why! The images for these databases and other information about them can be found in "A Proposal for Standard Graphics Environments," IEEE Computer Graphics & Applications, November 1987, pages 3-5. A corrected image of the tree appears in IEEE CG&A, January 1987, page 18. At present, the SPD package is available for the IBM PC on 360K 5.25" floppy for $5 from: MicroDoc, c/o F.W. Pospeschil, 3108 Jackson Street, Bellevue, Nebraska 68005. File Structure -------------- Six different procedural database generators are included. These were designed to span a fair range of primitives, modeling structures, lighting and surface conditions, background conditions, and other factors. A complexity factor is provided within each program to control the size of the database generated. This software package contains the following files: README - what you are now reading makefile - used to make the programs (in HP-UX Unix) def.h - some useful "C" definitions lib.h - globals and library routine declarations lib.c - library of routines balls.c - fractal ball object generator gears.c - 3D array of interlocking gears mountain.c - fractal mountain and 4 glass ball generator rings.c - pyramid of dodecahedral rings generator tetra.c - recursive tetrahedra generator tree.c - tree generator The compiled and linked programs will output a database in ASCII to stdout containing viewing parameters, lighting conditions, material properties, and geometric information. The data format is called the 'neutral file format' (or NFF) and is outlined in lib.c. This format is meant to be minimal and (hopefully) easy to attach to a user-written filter program which will convert the output into a file format of your choosing. Either of two sets of primitives can be selected for output. If OUTPUT_FORMAT is defined as OUTPUT_CURVES, the primitives are spheres, cones, cylinders, and polygons. If OUTPUT_FORMAT is set to OUTPUT_PATCHES, the primitives are simply polygonal patches and polygons (i.e. all other primitives are polygonalized). In this case OUTPUT_RESOLUTION is used to set the amount of polygonalization of non-polygonal primitives. In general, OUTPUT_CURVES is used for ray-trace timing tests, and OUTPUT_PATCHES for hidden surface and other polygon-based algorithm timings. SIZE_FACTOR is used to control the overall size of the database. Default values have been chosen such that the maximum number of primitives less than 10,000 is output. One purpose of SIZE_FACTOR is to avoid limiting the uses of these databases. Depending on the research being done and the computing facilities available, a larger or smaller number of primitives may be desired. SIZE_FACTOR can also be used to show how an algorithm's time changes as the complexity increases. Other parameters (for example, branching angles for "tree.c" and the fractal dimension in "mountain.c") are included for your own enjoyment, and so normally should not be changed if the database is used for timing tests. Note that light intensities, ambient components, etc. are not specified. These may be set however you prefer. The thrust of these databases is the testing of rendering speeds, and so the actual color should not affect these calculations. An ambient component should be used for all objects. A simple formula for an intensity for each light and the ambient component is the following: sqrt(# lights) / (# lights * 2). Database Analysis ----------------- The databases "mountain", and "tetra" consist of primitives of about the same size in a localized cluster, "balls" and "tree" are more varied clusters, and "rings" and "gears" are somewhat space-filling for the eye rays. Some other facts about these databases: balls gears mountain rings tetra tree ----- ----- -------- ----- ----- ---- primitives SP P PS YSP P OSP (where S=sphere, P=polygon, Y=cylinder, O=cone, in # of objects order) total prim. 7382 9345 8196 8401 4096 8191 polys 1417K 9345 8960 874K 4096 852K rend. polys 707K 4448 5275 435K 2496 426K ave. poly 0.26 173.6 53.2 2.95 35.4 0.09 vectors 4251K 55300 26880 2688K 12288 2621K rend. vectors 4251K 54692 26074 2688K 12288 2621K ave. vector 0.59 7.6 12.5 2.57 11.1 0.26 ave. # edges 3.00 5.92 3.00 3.08 3.00 3.08 lights 3 5 1 3 1 7 background 0% 7% 34% 0% 81% 35% specular yes yes yes yes no no transmitter no yes yes no no no eye hit rays 263169 245532 173422 263169 49806 170134 reflect rays 175616 305561 355355 316621 0 0 refract rays 0 208153 355355 0 0 0 shadow rays 954544 2126105 362657 1087366 46150 1099748 "total prim." is the total number of ray-tracing primitives (polygons, spheres, cylinders and cones) in the scene. The number of polygons and vectors generated is a function of the OUTPUT_RESOLUTION. The default value for this parameter is 4 for all databases. "polys" is the total number of polygons and polygonal patches generated when using OUTPUT_PATCHES. "rend. polys" is the number of polygons actually sent to the z-buffer (i.e. not culled and not fully clipped). "ave. poly" is the average rendered polygon size in pixels for a 512 x 512 resolution. Note that this size is the average of the number of pixels put on the screen by all unculled, not fully clipped polygons. Culled polygons are not counted, nor are pieces of polygons which are off-screen (clipped). For this statistic, all transparent objects are considered opaque. The area was calculated directly from the exact transformed vertices, not from the screen. Note that in practice there usually are more pixels per polygon actually rendered, mostly due to the polygon edges being fully rendered. It is because of this machine dependency that the purely geometric area is given. "vectors" is the total number of vectors generated. "rend. vector" is the number of vectors which were not fully clipped off the screen (note that no culling takes place in vector testing). "ave. vector" is similar to "ave. poly", being the average length of all rendered vectors in pixels. Note that culling is not performed for vectors. Again, the actual number of pixels rendered is machine dependent and could be different than this average vector length. "lights" is simply the number of lights in a scene. "background" is the percentage of background color (empty space) seen directly by the eye for the given view. It is calculated by "1 - ( eye hit rays/(513*513) )", since 513 x 513 rays are generated from the eye. "specular" tells if there are reflective objects in the scene, and "transmitter" if there are transparent objects. "eye hit rays" is the number of rays from the eye which actually hit an object (i.e. not the background). "reflect rays" is the total number of rays generated by reflection off of reflecting and transmitting surfaces. "refract rays" is the number of rays generated by transmitting surfaces. "shadow rays" is the sum total of rays shot towards the lights. Note that if a surface is facing away from a light, or the background is hit, a light ray is not formed. The numbers given can vary noticeably from a given ray tracer, but should all be within about 10%. "K" means exactly 1000 (not 1024), with number rounded to the nearest K. Testing Procedures ------------------ Below are listed the requirements for testing various algorithms. These test conditions should be realizable by most systems, and are meant to represent a common mode of operation for each algorithm. Special features which the hardware supports (or standard features which it lacks) should be noted with the statistics. Hardware Vector Testing: 1) Two separate tests should be performed. One test should be done at a resolution of 512 x 512. The second test should be done at the maximum square resolution representable on the screen (e.g. the HP SRX's resolution is 1280 x 1024, so run the tests at 1024 x 1024). The first test is done so that the same sized polygons are used in tests. The second test is done to test the display hardware using a more commonly used image resolution. 2) At least 24 bit planes should be used for color rendering, if available. If not, this should be noted and the best mapping mode available should be used. 3) Vectors should be flat-shaded, with no hidden line removal performed and no depth-cueing (but note these features if available). 4) The largest unit of display is the polygon. This means that no queueing of edge lists or optimization of move/draw commands can be performed by the user. All polygons must be rendered. This will mean that many edges will be drawn twice (since they are shared by polygons). 5) Timings will consist of only the time spent on the polygon draw calls, including the overhead of calling the display routine itself. This goal can most easily be realized by timing two versions of the testing program. One version reads the database and displays the results normally. The second version has all the system supplied vector drawing commands removed, and so will output no vectors. We can determine the time spent actually displaying vectors by taking the difference of these two timings. This time divided by the total number of vectors created by the database (listed in the table as "vectors") is the vector rate for the database. 6) Two separate rates should be calculated from the average rates computed. One is the display rate, which is the average of the gears, mountain, and tetra databases; the other is the throughput rate, which is the average of the balls, rings, and tree databases. The difference is that the databases used to calculate the display rate are more realistic as far as average polygon size presently used in hidden surface comparisons. This rate should better reflect the day-to-day performance expected of a machine. The databases for the throughput rate are characterized by very small polygons, and so will reflect the fastest speed of the hardware, as the effort to actually put the vectors on the screen should be minimal. Hardware Shaded Hidden Surface Testing: 1) As in hardware vector testing, two tests should be performed at the two resolutions. A display and throughput rate should be calculated. 2) All polygons in the databases are guaranteed to be planar. Polygons do not interpenetrate, and only in "rings.c" does cyclical overlapping occur. The first three points of a polygon always form an angle less than 180 degrees, and so can be used to compute the surface normal. This computation should be done by the hardware if possible (and if not, should be noted as such). 3) Polygons are one-sided for all databases (though transparent objects may have to be treated as two-sided), and so the hardware can use the surface normal to cull. It is not valid to cull (or perform any other clipping or shading function) using the software. 4) Polygons should be displayed using hidden surface rendering methods. No software sorting or other irregular manipulation of the data is allowed: the polygons should be rendered on a first-come first-served basis. 5) Up to seven light sources and an ambient contribution are used in the databases. If the hardware does not have this many, note it and test with the first few lights. Lights do not fall off with distance. Shadowing is not performed (though it's an incredible feature if it is). 6) Light sources are positional. If unavailable, assign the directional lights a vector given by the light position and the viewing "lookat" position. 7) Specular highlighting should be performed for surfaces with a specular component. The simple Phong distribution model is sufficient (though note improvements, if any). 8) Polygonal patches should have smooth shading, if available. In this case a Gouraud interpolation of the colors calculated at the vertices is the simplest algorithm allowed (again, note any more advanced algorithms). 9) Transparency should be used if available, and the technique explained. For example, the HP SRX uses "screen door" transparency, with a fill pattern of a checkerboard. Transparent objects are then rendered by filling every other pixel. 10) Overflow of color calculations is not guaranteed for these databases. As such, this should be handled by scaling to the maximum color contribution (e.g. r/g/b of 2 4 5 gets scaled to 0.4 0.8 1.0) if available. If not, simple clamping should be used and noted (e.g. an r/g/b of 0.8 1.1 1.5 is output as 0.8 1.0 1.0). 11) As in vector testing, timings should be done only for the actual polygon rendering routines. This means such costs as changing fill color, vertex format, or other state record information is not counted in the timing performed, but rather a flat-out speed test of the shaded polygon rate is made. Other Hardware Tests: A number of tests which could also be performed, but that I have ignored for now, include the following: 1) Flat-shaded rendering: render polygons using only their fill colors. Such renderings will look like mush, but would be useful for showing the difference in speed for a flat-shade vs. shaded rendering. 2) Hidden-line rendering: render flat shaded polygons the same color as the background, with edges a different color. Culling can be used to eliminate polygons. 3) No polypatch rendering: render polygonal patches but ignore their normal per vertex values (i.e. render them as simple flat polygons). Again, useful to show the time cost of using normals/vertex. Pointless, of course, for the databases without polygonal patches. 4) No specular rendering: render without specular highlighting. Useless for the databases without specular highlighting. 5) Total cull and clip rendering: render with the lookat direction reversed. Change the at vector to be equal to ( 2*from - at ) and up to be -up. No objects will be displayed. This test gives a sense of the speed of the cull and clip hardware. Such a full battery of tests would yield some interesting results. Hardware could be compared in a large number of categories, allowing users to be able to select equipment optimized for their application. I feel the most important rates are the vectors/second and the fully rendered polygons/second, as these two display methods bracket the gamut of realism offered by standard display hardware. Ray-Tracer Testing: 1) Assume the same conditions apply as in the shaded hidden surface testing outline, except as noted below. Note that the non-polygon (OUTPUT_CURVES) format should be used for ray-tracing tests. 2) All opaque (non-transmitting) primitives can be considered one-sided for rendering purposes, similar to how polygons were considered one- -sided for hidden surface testing. Only the outside of sphere, cone, and cylinder primitives are viewed in the databases. 3) Render at a resolution of 512 x 512 pixels, shooting rays at the corners (meaning that 513 x 513 eye rays will be created). The four corner contributions are averaged to arrive at a pixel value. If this is not done, note this fact. No pixel subdivision is performed. 4) The maximum tree depth is 5 (where the eye ray is of depth 1). 5) All rays hitting only specular and transmitting objects spawn reflection rays, unless the maximum ray depth was reached by the spawning ray. No adaptive tree depth cutoff is allowed; that is, all rays must be spawned (adaptive tree depth is a proven time saver and is also dependent on the color model used). 6) All rays hitting transmitting objects spawn refraction rays, unless the maximum ray depth was reached or total internal reflection occurs. Transmitting rays should be refracted using Snell's (i.e. should not pass straight through an object). If total internal reflection occurs, then only a reflection ray is generated at this node. 7) A shadow ray is not generated if the surface normal points away from the light. 8) Assume no hierarchy is given with the database (for example, color change cannot be used to signal a clustering). The ray tracing program itself can create its own hierarchy, of course. 9) Timing costs should be separated into at least two areas: preprocessing and ray-tracing. Preprocessing includes all time spent initializing, reading the database, and creating data structures needed to ray-trace. Preprocessing should be all the constant cost operations--those that do not change with the resolution of the image. Ray-tracing is the time actually spent tracing the rays (i.e. everything that is not preprocessing). 10) Other timing costs which would be of interest is a breakdown of times spent in the preprocessing and for actual ray-tracing. Examples include time spent creating a hierarchy, octree, or item buffer, and times spent on intersection the various primitives and on calculating the color. One major complaint with simple timing tests is that they tell little about the actual algorithm performance per se. One partial remedy to this problem is to also include in statistics the number of object intersection tests performed and the number of hits recorded. This information is useful for comparing techniques. One example would be comparing various automatic hierarchy algorithms by using these statistics. Other information should be included in ray-trace timing results that is often overlooked or ignored. Some basic information about the system configuration should be made: average speed of machine in terms of mips, Mflops, or Vaxen (i.e. the HP 320 is about 75 Kflops (single), which is about 30% of the speed of a VAX 11/780 FPA running vms 4.1, which is rated at 250 Kflops. However, an HP hardware engineer told me that the HP 320 is twice the speed of an 11/780 - it is unclear if this was incorrect. He might be right about overall machine speed in MIPS, for example); physical memory size (i.e. 6 Mbytes); special hardware such as a floating point accelerator or array processor, if different than the standard machine; system costs during tests (i.e. single user mode vs. time-share with other heavy users vs. serious operating system overhead); operating system (i.e. HP-UX version 5.2); and language used (i.e. optimized "C", single precision in general). Timings ------- Rendering time for test set on HP-320SRX: Vector (seconds) | Z-buffer (seconds) 512 x 512 1024 x 1024 | 512 x 512 1024 x 1024 balls 369.03 380.32 | 284.52 320.93 gears 6.19 5.26 | 6.84 4.82 mountain 4.83 4.37 | 3.84 2.88 rings 232.28 243.65 | 181.28 202.11 tetra 2.24 2.08 | 1.85 1.40 tree 226.54 236.44 | 176.17 196.68 Calculated performance of test set on HP-320SRX: Vector (vectors/second) | Z-buffer (polys/second) 512 x 512 1024 x 1024 | 512 x 512 1024 x 1024 gears 8937 10513 | 1365 1940 mountain 5563 6154 | 2331 3115 tetra 5496 5908 | 2216 2925 ------------------------------------------------------------------------------- Display average 6665 7525 | 1971 2660 (display average vector length for 512x512 is 10.4 pixels, 1024x1024 is 20.8) (display average polygon size for 512x512 is 87.4 pixels, 1024x1024 is 349.6) Vector (vectors/second) | Z-buffer (polys/second) 512 x 512 1024 x 1024 | 512 x 512 1024 x 1024 balls 11521 11179 | 4981 4416 rings 11572 11032 | 4819 4322 tree 11569 11084 | 4835 4331 ------------------------------------------------------------------------------- Throughput ave. 11554 11098 | 4878 4356 (throughput ave. vector length for 512x512 is 1.1 pixels, 1024x1024 is 2.3) (throughput ave. polygon size for 512x512 is 1.1 pixels, 1024x1024 is 4.4) The above times are computed from the average of 5 runs for each database. The timing increment is 0.02 seconds (50 Hz). Variance over the runs was quite small. Note that many of the tests at maximum square resolution (1024 x 1024) for the hidden surface timings are faster than those for the low resolution (512 x 512). The HP-320SRX is a workstation with 6 MByte memory and floating point accelerator. 24 bit color was used for display, and the full screen resolution is 1280 x 1024. Rendering time of the ray-traced test set on HP-320SRX (512 x 512 pixels): Setup Ray-Tracing | Polygon Sphere Cyl/Cone Bounding (hr:min:sec) | Tests Tests Tests Vol. Tests ------------------------------------------------------------------------------- balls 13:24 4:36:48 | 1084K 6253K 0 41219K gears 13:49 10:32:51 | 12908K 0 0 92071K mountain 11:46 2:59:58 | 3807K 4193K 0 31720K rings 31:56 13:30:40 | 1203K 5393K 15817K 94388K tetra 4:04 15:40 | 342K 0 0 1491K tree 7:37 3:43:23 | 1199K 2218K 973K 20258K A typical set of ray tracing intersection statistics for the tetra database is: [these statistics should be the same for all users] image size: 512 x 512 total number of pixels: 262144 [ 512 x 512 ] total number of trees generated: 263169 [ 513 x 513 ] total number of tree rays generated: 263169 [ no rays spawned ] number of eye rays which hit background: 213363 [ 81% ] average number of rays per tree: 1.000000 average number of rays per pixel: 1.003910 total number of shadow rays generated: 46150 [these tests vary depending on the ray-tracing algorithm used] Ray/Item Tests Performed: Polygonal: 342149 ( 55495 hit - 16.2 % ) Bounding Box: 1491181 ( 351754 hit - 23.6 % ) Total: 1833330 ( 407249 hit - 22.2 % ) Setup times are dominated by a verbose data format which causes a massive (usually tenfold) increase in size from the NFF file format to our in-house format. It should also be noted that for my sphere and bounding volume tests there are more tests than are strictly needed. This is because I require each light to be attached to an object (a sphere), which leads to extra testing in both of these categories. For what it's worth, my ray-tracer is based on hierarchical bounding boxes generated using Goldsmith & Salmon's automatic hierarchy method (see IEEE CG&A May 1987) and uses an item buffer (no light buffer yet) and shadow coherency. Something to think about is how the octree, SEADS, and other such algorithms perform when the background polygon dimensions are changed (thus changing the size of the outermost enclosing box, which changes the octree encoding of the environment). Same question with the effects on the algorithms of moving light source positions along the line defined by the light and "lookat" positions. Future Work ----------- These databases are not meant to be the ultimate in standards, but are presented as a first attempt at providing somewhat representative modelled environments. A number of extensions to the file format should be provided someday, along with new database generators which use them. The present databases do not contain holed polygons, spline patches, polygonal mesh or triangular strip data structures, or other proposed PHIGS+ extensions. Modeling matrices are not output, and CSG combined primitives are not included. As far as database geometry is concerned, most scenes have a preponderance of small primitives, and in general very few objects ever get clipped. If you find that these databases do not reflect the type of environments you render, please write and explain why (or better yet, write one or more programs that will generate your "typical" environments--maybe it will get put in the next release). Acknowledgements ---------------- I originally heard of this idea from Don Greenberg back in 1984. Some time earlier he and Ed Catmull had talked over coming up with some standard databases for testing algorithmic claims, and to them must go the credit for the basic concept. Many thanks to the reviewers, listed alphabetically: Kells Elmquist, Jeff Goldsmith, Donald Greenberg, Susan Spach, Rick Speer, K.R. Subramanian, John Wallace, and Louise Watson. Other people who have freely offered their ideas and opinions on this project include Brian Barsky, Andrew Glassner, Roy Hall, Chip Hatfield, Tim Kay, John Recker, Paul Strauss, and Chan Verbeck. These names are mentioned mostly as a listing of people interested in this idea. They do not necessarily agree (and in some cases strongly disagree) with the validity of the concept or the choice of databases. Your comments and suggestions on these databases are appreciated. Please send any timing results for software and hardware which you test.