Path: utzoo!utgpu!news-server.csri.toronto.edu!rutgers!apple!veritas!amdcad!mozart.amd.com!proton!tim
From: tim@proton.amd.com (Tim Olson)
Newsgroups: comp.benchmarks
Subject: Re: Sun weirdness.
Message-ID: <1990Dec18.144117.25312@mozart.amd.com>
Date: 18 Dec 90 14:41:17 GMT
References: <605@mephisto.edu>
Sender: usenet@mozart.amd.com (Usenet News)
Reply-To: tim@amd.com (Tim Olson)
Distribution: usa
Organization: Advanced Micro Devices, Austin, TX
Lines: 59

In article <605@mephisto.edu> juan@burdell.gatech.edu (Juan Orlandini) writes:
| 
| Can anyone tell me why the following program takes only .9 user seconds
| an .1 sys seconds on a sun 4/370 and 5.0 user seconds and 39.0 sys seconds
| on sparc 1's (1, 1+, IPC)?
| 
| ---------------------------- cut here -----------------------------
| #include <stdio.h>
| 
| main()
| {
|     int    num_times,row1,col2,col1,num_cols1,num_cols2,num_rows;
|     float  mat1[4][4],mat2[4][4],res[4][4];
|     float  temp;
|     
|     num_rows = 4;
|     num_cols1 = 4;
|     num_cols2 = 4;
|     for(num_times=0;num_times<10000;num_times++)
|     {
|       for(row1=0;row1<num_rows;row1++)
|       {
|         for(col2=0;col2<num_cols2;col2++)
|         {
|             temp = 0;
|             for(col1=0;col1<num_cols1;col1++)
|             {
|               temp = temp + mat1[row1][col1]*mat2[col1][col2];
|             }
|             res[row1][col2] = temp;
|         }
|     }
|       }
| }
|             

It's probably because the arrays mat1 and mat2 are local to the
function main(), but are never initialized.  Therefore, they can
contain garbage.  The large amount of system time you see is probably
the Sparcstation taking FP traps on invalid or denormalized inputs to
fix up the output per IEE-754.  I'm not sure about the 4/370, but it
may handle some of these in hardware, or it may just not have the same
random garbage on the stack.

If you move the declaration of mat1 and mat2 out of main, so they are
global and get initialized to zero (luckily, the bss zero is the same
as a single-precision 0.0, here), then you get closer to the time you
expect:

	1.6u 0.0s 0:01 100% 0+116k 4+0io 3pf+0w (on a Sparcstation 1)


This is a graphic example of why benchmarks should be self-checking,
or at least produce some output that can be checked by hand....

--
	-- Tim Olson
	Advanced Micro Devices
	(tim@amd.com)