Path: utzoo!utgpu!water!watmath!clyde!att!osu-cis!tut.cis.ohio-state.edu!mailrus!uwmcsd1!leah!bingvaxu!sunybcs!rutgers!okstate!mbc
From: mbc@a.cs.okstate.edu (Michael Carter)
Newsgroups: comp.graphics
Subject: Re: Vectorizing ray-object intersection calculations
Summary: ...object intersections aren't the problem ...
Keywords: ray tracing, vectorization
Message-ID: <3877@okstate.UUCP>
Date: 29 Aug 88 14:19:57 GMT
References: <594@fcs280s.ncifcrf.gov>
Organization: Oklahoma State Univ., Stillwater
Lines: 29

In article <594@fcs280s.ncifcrf.gov>, palmer@ncifcrf.gov (Thomas Palmer) writes:
> 
> Has anyone done any work on vectorizing ray-object intersection
> calculations?  The vectorizing C compiler on a Cray X/MP will not
> (automatically) vectorize loops smaller than about 5 or 6 iterations
> (nor should it - the vector register load time outweighs the gain).
> Typical ray tracing vector operations involve vectors of length 3 or 4.
> 
     The real problem in the inner ray tracing loop is not ray-object
intersections, but ray-bounding volume intersections.  If you refer
to the article by Kay and Kajiya from SIGGRAPH '86, they describe
a method of breaking down the OBJECT space into a hierarchical data
structure, and intersecting rays against simple bounding volumes
constructed from sets of planes.  This method queries the objects
in the order that they occur along the ray, therefore, NO SORTING IS
NEEDED.  It takes at lease three pairs of planes to completely
enclose an object, therefore this intersection calculation could be
done efficiently, in parallel (on perhaps more than one object at a
time!) on a vector machine.  Most of the time is spent in this ray-
bounding volume intersection loop, and not in the ray-object intersection
algorithms.
     I realize that this is not something that the C compiler can do
on its own, but remember: no pain -- no gain.  (-:

-- 
Michael B. Carter
Department of Electrical and Computer Engineering, Oklahoma State University
UUCP:  {cbosgd, ea, ihnp4, isucs1, mcvax, pesnta, uokvax}!okstate!mbc
ARPA:  mbc%okstate.csnet@csnet-relay.arpa