Re[2]: gcc compiler flags

From: Jason J Ding (Jason_J_Ding@ccm.sc.intel.com)
Date: Fri Oct 13 1995 - 11:41:00 CDT


Text item:

        Hello Russell,

        Thanks for sending me the gcc compiler flags.
        I was going to get the gcc numbers and see the differences
        between gcc and Intel Reference Compiler. However,
        my hard disk went down after I collected P5-133/and P5-90
        numbers (I moved disk around too much?). Now I have to
        re-install the UNIXware system.

        The SAXPYing bandwidth number I got from Gateway 2000 P5-133
        is about 33% higher than that you reported to John. I used
        256 PBSRAM L2 cache, EDO DRAM, UNIXware 2.0, and Intel Reference
        Compiler. My compiler flags are: "icc -tp p5 -ip stream_d.c".
        
        I also noticed there is a Revision 2.1 (August 30, 1995) and
        Revision 2.0 dated Sept 30, 1991 (?). However, it was stated
        that "modified slightly to account for timing overhead .. by
        R. L. Carter 5/4/95". So I assume it was the Sept 30, 1995
        version. But it was numbered Revision 2.0?

        Anyway, my results were from Revision 2.0 (I also noticed the
        format of your results to John was from that version).

        --jason

        

Hello Jason,
I used "gcc-2.6.3 -O4 -m486 -funroll-loops -ffast-math".
Note that there is no Pentium optimization. I have
a version of the 2.6.3 compiler with the original gcc-2.0
Intel diffs folded in (more or less), but the higher levels
of optimization that access the Pentium specific optimizations
are too buggy to use. A pretty big chunk of people, namely
the BSDI, Linux, FreeBSD, and NetBSD users, would benefit
a lot from having improved Pentium and Pentium Pro optimization
folded into gcc. As it is the performance is about 30%
less on some high profile loops. On things that are
main memory bandwidth limited there isn't much difference.

How much faster is the reference compiler on the stream
benchmark? Also note that I'm using a slightly different
version of stream_d.c that accurately computes the rates,
and they are slightly slower. It's available at

http://www.geli.com/src/cstream.tar.gz

Thanks,
Russell Carter
Geli Engineering
rcarter@geli.com
(408) 246-6905

Text item: External Message Header

The following mail header is for administrative use
and may be ignored unless there are problems.

***IF THERE ARE PROBLEMS SAVE THESE HEADERS***.

Cc: rcarter@geli.com
Subject: Re: gcc compiler flags
To: Jason_J_Ding@ccm.sc.intel.com
Message-Id: <199510121745.KAA05040@geli.clusternet>
From: "Russell L. Carter" <rcarter@geli.com>
Date: Thu, 12 Oct 1995 10:45:46 -0700
Received: (from rcarter@localhost) by geli.clusternet (8.6.12/8.6.9) id KAA05040
; Thu, 12 Oct 1995 10:45:46 -0700
Received: from geli.clusternet (rcarter.vip.best.com [204.156.137.2]) by blob.be
st.net (8.6.12/8.6.5) with ESMTP id KAA12548 for <Jason_J_Ding@ccm.sc.intel.com>
; Thu, 12 Oct 1995 10:48:17 -0700
Received: from blob.best.net by aurora.intel.com (5.65/10.0i); Thu, 12 Oct 95 10
:52:07 -0700
Received: from aurora.intel.com by ccmgate1.sc.intel.com with smtp
     (Smail3.1.28.1 #3) id m0t3Rjv-000S0WC; Thu, 12 Oct 95 10:48 WET DST



This archive was generated by hypermail 2b29 : Tue Apr 18 2000 - 05:23:05 CDT