You are right!!! I got the same idea the moment I hit return on my
previous post. So I increased the loop counter and used a stopwatch
(external and crude) to measure the time and it pretty much takes the same
amount of time with interrupts turned on or off.
So I am back to square one. The floating point operations whether I use
kernel emulation or -msoft-lib are too slow for my use.
Thanks to all those who looked at the email.
--
Pawan Singh
[EMAIL PROTECTED]
650-404-0213
On Tue, 27 Jun 2000, Nicolas Pitre wrote:
>
>
> On Tue, 27 Jun 2000, Pawan Singh wrote:
>
> > 1. The piece of code I am running is a user process.
> > 2. Only reason I tried CLI() and STI() is so that I can find out how much
> > time does the CPU take to execute the code - just as a benchmark. I will
> > never use cli and sti in the user land.
> >
> > 3. I finally have one more clue. It does not matter if it is a floating
> > point operation inside the loop. One gets the same results with integer
> > operations. The result gets worse and worse depending on the number of
> > instructions in the loop.
> >
> > 4. I have made sure that compiler has not optimized anything. The piece of
> > code I typed was just a quick demonstration. Add to the code something after
> > the loop (like printf) which uses the sum total being produced by the loop.
> >
> > So in short it seems to me is this - for one integer add:
> > 1. Interrupts turned off: 1000000 integer additions take 10ms.
> > 2. With interrupts turned on 10000000 integer additions take 270ms.
> >
> > To me it is 27 times kernel overhead.
> >
> > If I increase the number of integer instructions in the loop to what
> > floating point code does for two float additions i.e. 100 instructions,
> > 1. Interrupts turned off: 10000 loop interations take 10ms.
> > 2. With interrupts turned on 100000 loop iterations take 1000 to 2000
> > seconds.
> >
> > To me in this case it is 100-200 times kernel overhead.
> >
> > These numbers can be reproduced on any 486, 386 with 8K L1 cache.
> >
> > Comments?
>
> How do you measure the elapsed time?
>
> Since time counters in the kernel are 10 ms precision and incremented at
> each timer interrupt, you just don't have the real elapsed time when IRQs
> are masqued.
>
>
> Nicolas
>
--
To unsubscribe from this list, send a message to [EMAIL PROTECTED]
with the command "unsubscribe linux-embedded" in the message body.
For more information, see <http://waste.org/mail/linux-embedded>.
