Path: chuka.playstation.co.uk!news
From: sceetech <ps_yaroze@interactive.sony.com>
Newsgroups: scee.yaroze.programming.libraries
Subject: Re: Rotation Matrix Functions
Date: Mon, 07 Apr 1997 12:24:29 +0100
Organization: SCEE
Lines: 52
Message-ID: <3348D96D.289@interactive.sony.com>
References: <01bc4074$bd79ee80$d603e8c3@fred> <334A0C83.5959@micronetics.com>
Reply-To: ps_yaroze@interactive.sony.com
NNTP-Posting-Host: 194.203.13.10
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
X-Mailer: Mozilla 3.01 (Win95; I)

Jim wrote:
> 
> Hi Mike,
> 
> I've also seen this.
> 
> Are you directly additive modifying the matrix (coord.m i think) element
> of GsCOORDINATE2 for each frame?  I think your objects are scaling
> because  of a small error creeping in for each rotation operation.
> Because you are modifying the same matrix for every frame the error get
> bigger and bigger until every thing starts freaking out. :-)
> 
> If you take a copy of the initial matrix each frame and modify that with
> the rotation metrics you should be ok.  I don't think this will apply to
> translation.
> 
> For example rather than rotating an object 50 degrees on frame one and
> by another 50 the next,
> First rotate by 50 from the initial matrix, then by 100 again from the
> initial for the next frame.
> This will stop the error multiplying into one huge climax! :-)
> 
> This may make no sense just yet because its monday morning before
> coffee.  I'll reread this this afternoon and post an errata then. :-)
> 
> Regards
> 
> Jim
> 

It is right that cumulative changes to a matrix are liable
to progressively distort it due to accumulating errors in
integer approximations; these distortions will have the
effect of scaling and shearing 3d polygon models.

Solutions to this:
(i) if only ever rotating a model in one plane, can just store
a single angle on that plane; each frame, regenerate the matrix
from the original matrix and the angle, thereby ensuring
minimal and constant error
(ii) if rotating in more than one plane, this approach cannot
be used because the order of rotations matters; building and
using an ordered list of rotations quickly becomes too
inefficient and cumbersome. One way to deal with this is
to monitor the error: eg every N frames, inspect the matrix,
look at it as 3 vectors, one per axis of the the object;
usually each started off life with size ONE (ie 4096),
hence test the actual sizes and rescale to ONE if
necessary. Les than ideal, but does prevent runaway errors.


	Lewis