Xref: utzoo comp.benchmarks:489 comp.graphics:16920
Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!usc!elroy.jpl.nasa.gov!decwrl!sgi!shinobu!odin!bruceh
From: bruceh@sgi.com (Bruce R. Holloway)
Newsgroups: comp.benchmarks,comp.graphics
Subject: Re: Truth or Dare?
Message-ID: <1991Mar29.005630.22949@odin.corp.sgi.com>
Date: 29 Mar 91 00:56:30 GMT
References: <1991Mar27.174104.26867@hellgate.utah.edu>
Sender: news@odin.corp.sgi.com (Net News)
Followup-To: comp.benchmarks
Distribution: comp
Organization: Silicon Graphics, Inc., Mountain View, CA
Lines: 55

In article <1991Mar27.174104.26867@hellgate.utah.edu> thomson@cs.utah.edu (Rich Thomson) writes:
>Several times on comp.graphics the SGI folks have offered up the
>famous "1 million polygons/second" quotation from marketing literature
>when referring to peak performance on the VGX series of machines.
>
>The polygons in question are 50 pixel triangle strips that are flat
>shaded.  Also, they may have rendered them without the Z buffer turned
>on (the spec sheet is away from me at the moment).
>...
>So: is this a "theoretical" number, or is it attainable by a program,
>no matter how contorted?

We have a suite of performance measurement programs which allow lots of
primitives with different characteristics to be tried.  I ran a certain
case and got the following output:

* Mesh, DisplayList, Subpixel, RGBmode, Flat, Area = 50.000000, PYM_FILL, MeshesPerSecond = 886681

My machine is only a 5-span, which means I have just one raster memory board.
Apparently the quoted performance numbers are for a 10-span system.  With
slightly smaller triangles I obtained:

* Mesh, DisplayList, Subpixel, RGBmode, Flat, Area = 40.500000, PYM_FILL, MeshesPerSecond = 1022706

The inner loops looked like this:

    makeobj(*name=genobj());

    for(jdx=0; jdx<OBJSZ; jdx++) {
	fp=(float *)rectbuf;
        bgntmesh();
	  v3f(fp); v3f(fp+4);
	  for (fp+=8, i=LoopCount; i>0; --i,fp+=40) {
                 v3f(fp);    v3f(fp+4);
                 v3f(fp+8);  v3f(fp+12);
                 v3f(fp+16); v3f(fp+20);
                 v3f(fp+24); v3f(fp+28);
                 v3f(fp+32); v3f(fp+36);
	  }
        endtmesh();
    }

    closeobj();

    grestartwatch();
    for (i=events/(objects*OBJSZ); i>0; i--)
       callobj(Obj);
    sec = gstopwatch();
    return((int) events/sec);

My cursory inspection of the program revealed zbuffer(TRUE) &
zfunction(ZF_ALWAYS).  I'd send something simple & complete if I had time,
but this is the best I can do right now.

Regards, bruceh