Click here to get portable GPL source code and Windows binaries.
Benchmarking results from 100s of systems
Benchmarking results from the renderer's execution on 100s of systems (Phoronix).
One of my hobbies that has persisted over the years is my real-time
pure-software 3D renderer. I began writing it in the days of Hercules
and CGA cards... and will probably be playing with the code for as long as I live :‑)
My priorities, ever since I started doing this, are simple: try to make the code as clear and concise as possible,
while using good algorithms to improve the rendering speed. In plain words, my primary care is the clarity
of the code - as well as the renderer's speed.
Generated in the renderer's raytracing mode, with reflections and refractions enabled.
Conciseness and clarity are mostly accomplished via C++ templates, that unify the incremental calculations for the rasterizers
and the ray intersections for the raytracers. And as for speed, we are now firmly in the age of multi-core CPUs - so
software rasterizing can (finally) do per-pixel lighting and soft-shadows in real-time, while raytracing can generate
beautiful images in a matter of seconds.
The renderer allows switching between rasterizing and raytracing in real-time - so one navigates quickly with the
rasterizers, and requests a raytracing when the viewpoint is interesting.
On a 4-core Phenom, the shadow-mapped version runs at 60 frames/sec. The raytraced version: 3 sec/frame ;‑)
This is a (more or less) clean implementation of the
basic algorithms in polygon-based 3D graphics. The code includes...
3D transformations (from world coordinates, to camera-space/light-space coordinates)
Point rendering (vertex-based or triangle-based)
Anti-aliased lines rendering
Gouraud shading (complete Phong equation calculated per vertex)
Phong shading (complete Phong equation calculated per pixel)
Z-Buffer hidden surface removal
Soft shadow mapping
Raytracing, with shadows, reflections, refractions, ambient occlusion and anti-aliasing
Portable display and keyboard handling through libSDL
The supported 3D formats are:
.3ds, i.e. the well known 3D Studio format (via lib3ds)
.tri, a simple binary dump of vertex and triangle data
Is orchestrated via autoconf/automake, so it will
compile and run cleanly on most platforms (tested so far on Linux/x86,
Mac OS/X, Windows (using TDM/MinGW gcc), OpenSolaris (GCC/TBB), OpenBSD/amd64
Includes a separate VisualC directory for Windows/MSVC users,
with all dependencies pre-packaged for easy compilation
Can be configured to use
OpenMP, provided that your compiler's
support for OpenMP is mature enough (e.g. GCC since version 4.3.2)
Uses C++ template-based metaprogramming, in order
to move as much rendering logic as possible from run-time to compile-time.
This is a software-only renderer, so don't expect hardware
class (OpenGL) speeds. Then again, speed is a relative thing: the train
object (available inside the source package, in the "3D-Objects" folder)
was rendered (in soft-shadows mode) at a meager 6fps on an Athlon
XP, back in 2003. Around 2005, however, a Pentium4 desktop at work took
this up to 11 fps. As of 2007, by way of Intel's Threading Building Blocks
(or OpenMP) the code uses both cores of a Core2Duo to run at 23fps...
And since it uses TBB/OpenMP, it will automatically make use of any
additional cores... so give the CPUs a few more years... :‑)
Update, November 2009: On a 4-core AMD Phenom at 3.2GHz, the
train now spins at 80 frames per second... Give me more cores! :‑) Update, September 2017: On a 16-core Intel Core i9 7960X...
718 frames per second! Update, June 2018: Phoronix shows the evolution of my renderer's
speed across 28 CPUs...
The code also runs 20-25% faster if compiled under 64-bit environments.
Skipping points rendering, lets begin with an anti-aliased rendering of a train:
Notice the anti-aliasing in the zoomed section
Moving on to more interesting rendering modes: the same train looks far better
with shadow mapping, which allows rendering self-shadowing objects in real-time,
even when using multiple light sources (35 frames per second on a Core2Duo):
Shadow mapping - self shadowing
In many such 3D models self-shadowing is easily identifiable, especially if we
This nice chessboard exhibits it, too - running with shadow-mapping at 40 fps on a Core2Duo:
Gotta find time to hook this up with gnuchess! :‑)
The renderer also includes a SAH/AABB/BVH raytracer (i.e. a raytracer
that accelerates ray intersections using a Bounded Volume Hierarchy,
formed from Axis-Aligned Bounding Boxes, created via a Surface Area
Heuristic). The results are of course, not real-time - but
quality-wise, they are on a class of their own: here's how the chessboard
looks when raytraced with reflections and
shadows (11 seconds to render this frame on the same machine):
Raytracing is orders of magnitude slower than rasterizing, but it creates beautiful images...
I recently (Feb 2011) completed my first steps in speeding it up with CUDA (check it out).
Notice also how zooming-in on a pawn shows reflections of reflections...
...and finally, this is how refractions create a glass dragon (31 seconds to render on a Core2Duo):
Which brings us to the matter of speed, again: The program allows changing the rendering mode at runtime, and
therefore offers interactive control of the balance between rendering speed and rendering quality:
Shadow mapping and soft shadows
Shadow mapping offers a good balance between rendering quality and rendering speed.
Shadow maps are "special pictures" rendered along the normal rendering pipeline,
but from the point of view of the light source. They provide the "light-height"
information that tells the rasterizer when the pixel drawn is in shadow and when
not. In case you were wondering, here is what a
looks like (67KB image).
Normally, shadow maps generate sharp, "pixelated" shadow edges, because of the
sampling of the shadow map. To improve this, instead of sampling only one "shadow pixel",
the renderer can also use a weighted average of its neighbours, and thus provide
nice looking soft-shadows in real-time:
Self-shadowing at its finest...
Fast though it is, shadow-mapping has an issue if you zoom-in: the artifacts of the shadowmap sampling become annoying... In "deep" zooms, the renderer can be switched (at runtime) to raytracing mode, to create the correct shadows:
The "raycasted shadows" mode that I implemented in late 2010, was offering a compromise between the speed impact of a raytracer and the rendering artifacts of a rasterizer: it gave shadows the quality of a raytracer, but maintained some of the speed of a rasterizer, since everything else except the shadows was done via screen-space-linear interpolations: projected screen space coordinates, normal vectors used in Phong lighting, Z-Buffer handling, etc. It was also ported to CUDA, and got a hefty speedup.
Two weeks later, I removed this mode in favour of a full raytracer - it was slower than the rasterizer modes anyway, and a full raytracer offers far better quality. It still exists in the CUDA port, if you are interested.
The renderer uses ambient occlusion to significantly improve the rendering
quality of indirectly lit areas.
For the rasterizers, it linearly interpolates
(per-pixel) the ambient occlusion coefficient, which must be pre-calculated
per vertex and stored in the model (see below, "Creating more 3D objects on
For the raytracer, by uncommenting
the #define AMBIENT_OCCLUSION, you will enable a stochastic
ambient occlusion calculation for each raytraced pixel:
When a triangle is intersected by a primary ray, AMBIENT_SAMPLES
rays will be spawned from the intersection point, and they will be used to
calculate the ratio of ambient light at that point.
The difference is very clear:
With and without ambient occlusion
Here's an ambient-occlusion raytracing of Sponza (with 32 ambient rays cast per pixel):
...and another of a conference room:
Download, compile and run
Warning about hyper-threading
If your CPU uses hyper-threading, your "virtual" cores may help or may destroy
your performance. On an Atom 330 (2 real cores, each one appearing as two "virtual"),
the "virtual" cores help a lot: running with four threads, the raytracer is 1.3x faster
than running with two. On an dual-CPU, Intel Xeon W5580 however (total of 8 real cores,
appearing as 16 "virtual"), the speed increases almost linearly as we increase threads,
until we reach 8 - and then the speed nose-dives, with the 16 thread version being
63 times slower (!).
If your CPU has hyper-threading enabled, make sure you check the runtime performance of the
renderer by exercising direct control over the number of threads: use the OMP_NUM_THREADS
environment variable to control how many threads OpenMP uses.
Just open the project solution (under VisualC/) and compile for
Release mode. It is configured by default to use Intel TBB for multithreading,
since Microsoft decided to omit OpenMP support from the free version
of its compiler (the Visual C++ Express Edition). All dependencies (include files
and libraries for SDL and TBB) are pre-packaged under VisualC/, so compilation is
as easy as it can get.
When the binary is built, right-click on
"Renderer-2.x" in the Solution explorer, and select "Properties".
Click on "Configuration Properties/Debugging", and enter
inside the "Command Arguments" text box. Click on OK, hit Ctrl-F5,
and you should be seeing the chessboard spinning. Use the controls
described below to fly around the object.
The default compilation options are set for maximum optimization, using
If you have the commercial version of the compiler (which supports
OpenMP) you can switch from TBB to OpenMP:
Configuration Properties - C/C++ - Language - OpenMP: Set To "Yes"
Configuration Properties - Preprocessor - Definitions: Change USE_TBB to USE_OPENMP
For everybody else (Linux, BSDs, Mac OS/X, etc)
Compilation follows the well known procedure...
Voyager (yep, I Gimp-ed the lens flare :‑)
The source package includes a copy of the sources for lib3ds 1.3.0, and the
build process will automatically build lib3ds first.
SSE, SSE2 and SSSE3 x86 SIMD optimizations will be detected by configure,
but if you have a non-Intel CPU, pass your own CXXFLAGS flags, e.g.
bash$ CXXFLAGS="-maltivec" ./configure
Compiling under 64-bit environments (e.g. AMD64 or Intel EM64T) is further improving speed;
compiled with the same options, the code runs 25% faster under my 64-bit Debian.
A note for Mac OS/X and FreeBSD developers: The default
FreeBSD and Mac OS/X environments (XCode) include
an old version of GCC (4.2.x). This version is known to have issues with OpenMP,
so if you do use it, your only available option with multicore machines is
Intel TBB (which works fine). You can, however, download the latest GCC from
ports, if you use FreeBSD, or from
High Performance Computing for Mac OS/X -
they both offer the latest GCC series. Results are much better this way: OpenMP works fine,
and support for the SSE-based -mrecip option boosts the speed by more than 30%.
Some of the Parthenon's art (Elgin marbles)...
After a successful make, fly around the objects with:
bash$ cd 3D-Objects
bash$ ../src/renderer/renderer chessboard.tri
Hit 'R' to stop/start auto-spin.
Use the cursor keys, 'A' and 'Z' to pilot.
Rotate the light with 'W', 'Q'.
'S' and 'F' are 'strafe' left/right, 'E' and 'D' are 'strafe' up/down.
(strafe keys don't work in auto-spin mode).
The '1' to '0' keys as well as the PageUp/PageDown change the rendering mode, cycling through:
Points via triangle culling
Ambient (when ambient occlusion data are available in the 3D model, this actually looks good)
Gouraud (complete Phong lighting per vertex)
Phong (complete Phong lighting per pixel)
Phong and shadow maps
Phong and soft shadow maps
Raytracing, with shadows and reflections
Raytracing, with shadows, reflections and anti-aliasing
Try the other 3D objects, too: trainColor.tri, legocar.3ds, pharaoh.ply, etc...
Command line parameters
Usage: renderer [OPTIONS] [FILENAME]
-h this help
-r print FPS reports to stdout (every 5 seconds)
-b benchmark rendering of N frames (default: 100)
-n N set number of benchmarking frames
-w use two lights
-m <mode> rendering mode:
1 : point mode
2 : points based on triangles (culling,color)
3 : triangles, wireframe anti-aliased
4 : triangles, ambient colors
5 : triangles, Gouraud shading, ZBuffer
6 : triangles, per-pixel Phong, ZBuffer
7 : triangles, per-pixel Phong, ZBuffer, Shadowmaps
8 : triangles, per-pixel Phong, ZBuffer, Soft shadowmaps
9 : triangles, per-pixel Phong, ZBuffer, raycasted shadows
0 : raytracing, with shadows, reflections and anti-aliasing
Creating more 3D objects on your own
The rasterizer output is looking much better if the model carries pre-calculated
ambient occlusion information per vertex. To do this:
Load it up in shadevis and hit ENTER
to have shadevis calculate the ambient occlusion factors per vertex. After that,
hit 'D' as many times as necessary to lower the diffuse light to 0%, and hit 'a'
to pump up the ambient to 100%. Hit 'S' to save the object.
Load the saved '..._vis.ply' with my renderer.
Update: You can now use Meshlab to perform the same work, from:
"Filters / Color creation and processing / Ambient occlusion per vertex". Then
save your object as .ply (Stanford polygon format) and make sure you
have only "Color" (from the "Vert" group) selected (also, uncheck "Binary encoding").
The dark side... of coding SMP
Rant 1: Why did you do this, you crazy person?
Well... I've always loved coding real-time 3D graphics. Experimenting
with new algorithms, trying to make things run faster, look better...
And as a side effect, I became a better coder :‑)
Anyway, these sources are
my "reference" implementations. At some point around 2003, I decided that it
was time to clean up the code that I've been hacking on over the years
and focus on code clarity - ignoring execution speed. To that end, floating
point is used almost everywhere (fixed-point begone!) and this being Phong
shading, the complete lighting equation is calculated per pixel. I basically
created a "clean" implementation of everything I have ever learned about
polygon-related graphics. The clarity of the code also paved the way for the
OpenGL and CUDA versions...
Rant 2: Tales of Multicore
This code was single threaded until late 2007. At that point, I heard
about OpenMP, and decided to try it out. I was amazed at how easy it was
to make the code "OpenMP-aware": I simply added a couple of pragmas in the
for-loops that drew the triangles and the shadow buffers, and ...presto!
The only things I had to change were static variables, which had to be
moved to stack space. Threading code can't tolerate global/static data,
because race conditions immediately appeared when more than one thread
worked on them.
Once I began using OpenMP, the
renderer crashed many C++ compilers.
As of 2009, they have finally adapted!
Only two compilers truly supported OpenMP at the time: Intel's compiler
(version 8.1) and Microsoft's CL. GCC unfortunately died with
'internal compiler error'. I reported this to the GCC forums, found out
that I was not the only one who had noticed, and was told (by the forum
guys) to wait.
While waiting for GCC to catch up, I kept researching multicore technologies.
Functional languages seem particularly adept to SMP, and I've put them
next in line in my R&D agenda (Ocaml and F# in particular). Before leaving
C++ behind, though, I heard about Intel Threading Building Blocks (TBB)
and decided to put them to the test. TBB is a portable set of C++ templates
that makes writing threading code a lot easier than legacy APIs
(CreateThread, _beginthread, pthread_create, etc).
TBB is also open-source, so it was easy to work with it and figure out its
internals. Truth be told, it also required more changes in my code
(OpenMP required almost none). Still, it is a vast improvement compared
to conventional threading APIs.
I must also confess that I have not invested a lot of effort in using these
technologies; I only enhanced two of my main rendering loops to make them SMP
aware. Still, this was enough to boost the speed (on a Core2Duo) by 80%!
Judging by the gain/effort ratio, this is one of the best bargains I've ever
As of now (October 2008), GCC 4.3.2 is up to speed and compiles OpenMP
code just fine. TBB is of course running perfectly (since it is simply a C++
template library), so choose freely between any of the two, and easily
achieve portable multithreading.
When I say portable, I mean it: these are the tests I did...
Poor OpenBSD4.3/64: it doesn't have real, SMP threads. Not yet, at least :‑)
It only has user-space ones (as Linux did at some point).
But it does compile the code, albeit in single-threaded mode.
Talk about portable code!
Dynamic scheduling makes sure all cores are kept busy, even for low tesselations
If you're still in the... dark ages and use legacy APIs (CreateThread,
_beginthread, pthread_create, etc) you are really missing out:
Under both OpenMP and Intel TBB, I increased the rendering frame rate of
the train object by more than 40%, by simply replacing...
#pragma omp parallel for
#pragma omp parallel for schedule(dynamic,100)
(similar change for TBB, at code inside Scene.cc).
Why? Because these
modern threading APIs allow us to easily adapt to different loads per thread, by
using dynamic thread scheduling.
interested in reading the comments or leaving a comment behind please try again with a
different browser or from a different connection.