[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: glibc on ARM Processor
From: |
Andreas Jaeger |
Subject: |
Re: glibc on ARM Processor |
Date: |
02 May 2001 07:48:33 +0200 |
User-agent: |
Gnus/5.090003 (Oort Gnus v0.03) XEmacs/21.1 (Channel Islands) |
"Robert Couse" <address@hidden> writes:
> Mr. Jaeger,
> We are currently working on some C++ software that runs under VxWorks and we
> will be eventually porting it to Linux. The project is an embedded ARM 7
> network IC and we are doing our development work on the ARM Integrator Board
> with a ARM7TDMI. Both VxWorks/Tornado ( WindRiver ) and, of course, Linux
> uses the GNU compiler and the glibc library. We ran into some timing issues
> and, after an investigation, we found out that the issues are in the
> floating point library. We created a benchmark and compared it to our
> previous platform - i960 processor and with the library that is shipped by
> ARM in their Developer Suite. We have also compared it to Newlib with
> comparable results.
>
> Here is a comparison of running the attached sample code:
>
> Platform Tool Vendor Loop count Time (seconds)
> -----------------------------------------------------
> i960jx 33mhz Tornado (gnu) 100,000 8
> ARM7TDMI 20 Mhz Tornado (gnu) 1,000 13
> ARM7TDMI 20 Mhz ARM 100,000 23
Can you interpret those values for me - which one is using glibc
(which version)?
> My questions are:
> 1. Are you aware of the speed differences between the ARM and glibc speeds?
I personally not but I hope others will chime in.
> 2. Are there any glibc options or new releases which will speed up the
> library?
Which glibc version are you using?
> 3. Do you know of any alternatives to glibc which are faster?
No.
Please note that your test uses float variables for the test and uses:
val = (10*log10(power/(fsc * fsc))-(46*3.01)-3.65-powerCutBack +
initGain - gain);
Please use log10f so that the calculations is done in float and not in
double. Does this help?
Andreas
--
Andreas Jaeger
SuSE Labs address@hidden
private address@hidden
http://www.suse.de/~aj