GENERAL INFORMATION:
This code computes the equilibrium distribution of light in a scene
using the hierarchical diffuse radiosity method. A description of the
sequential hierarchical radiosity method can be found in
Pat Hanrahan, David Salzman and Larry Aupperle, "A Rapid Hierarchical
Radiosity Algorithm", Proc. SIGGRAPH 1991.
Descriptions of the parallel algorithm can be found in
Jaswinder Pal Singh, Anoop Gupta and Marc Levoy, "Parallel
Visualization Algorithms: Performance and Architectural Implications",
IEEE Computer, July 1994.
or in
Jaswinder Pal Singh, et al, "Load Balancing and Data Locality in
Hierarchical N-body Methods", Stanford Univ. Tech Report CSL-TR-92-505
(to appear in JPDC).
A detailed description will also be in the SPLASH-2 report.
The parallelism is managed with distributed task queues and task
stealing, and there is one task queue per processor.
RUNNING THE PROGRAM:
To see how to run the program, please see the comment at the top
of the main.C file or run it as "RADIOSITY -h". Many of the
command-line options control the accuracy of the program's computation
and have default values set in the program (which can be overridden by
command-line specifications). These are indicated as such in the
comments or usage statement, and we recommend that they be left at
their default values for base SPLASH-2 runs.
The program runs in different modes, either interactive (using the SGI
GL library) or batch. The batch mode does not attempt to display the
rendered image, while the interactive mode brings up a GL window with
knobs and dials to set parameters and run the program. If you are
running on a system that does not support GL or you do not want to
display the resulting scene, please use the -batch option on the
command line. The makefile shows you how to link the GL libraries.
This sample makefile is for a machine on which we do not want to
display, in which case we use the glibdumb and glibps libraries, which
do not support displaying and are provided with the code. The only
real way to verify the correctness of the program is to view the
result using GL. The commented lines in the print_statistics routine
can be uncommented to print several runtime statistics, but these are
nondeterministic in parallel since the program does not follow a
deterministic execution path and does not arrive at exactly the same
result in different parallel runs (the radiosity algorithm is
iterative to convergence, and even the path is nondeterministic).
The way the program is written, it compiles in the input description
of the scene in polygon coordinates. This is in the file room_model.C,
which contains the descriptions of two scenes. One is the room scene
originally used by Hanrahan et al in their SIGGRAPH paper, and the
other is an artificial extension of this scene (removing a wall of the
room and introducing some of the features of this room in the
neighboring room thus created). To use the former, use the -room
command line option; for the latter, use -largeroom. -room is what we
call the base SPLASH problem. If you don't specify -room or
-largeroom, the program will default to a small dummy test scene which
is useful only for debugging the program and verifying that it works.
There are two types of compile-time flags that can be set in the code.
One controls the manner in which patches are assigned to processors at
the beginning of each time-step in the radiosity iteration. The simple
case is to assign the same statically chosen set of patches to the
same processor in every iteration (and rely completely on task
stealing for load balancing); this is what happens by default, or if
-DPATCH_ASSIGNMENT_STATIC is used as a compile-time flag. The more
sophisticated case is to do an assignment at the beginning of a time
step based on the costs of patches profiled in the previous time step;
this is what happens if the -DPATCH_ASSIGNMENT_COSTBASED flag is used
at compile-time. The latter can yield less stealing, but uses more
synchronization to keep track of costs and hence has more overhead.
We recommend using the default (not defining anything, so static is
used) in the base SPLASH runs.
The other type of compile-time flag special cases some machines in
terms of maximum number of processors etc. (grep for #if in *.C and
*.H to find these) and even memory consistency model in one instance.
BASE PROBLEM SIZE:
The base problem size we recommend is to use the program as follows:
RADIOSITY -p ? -batch -room
where ? is the number of processors. This sets the ae, bf etc flags
(see comment at top of rad_main.C or run "RADIOSITY -h" to see what
these are) to their default values. The default values can be found
by looking at the comment at the top of rad_main.C or running
"RADIOSITY -h".
DATA DISTRIBUTION:
Data distribution is very difficult in this code, except for some
per-process data structures (typically arrays of structures indexed by
process_id; grep for MAX_PROCESSORS to find them).
General data distribution of other (scene, e.g.), however, does not
make much difference to performance on the Stanford DASH
multiprocessor.
Sanchayan Maity
0f4b39775c