[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

21164 speed definitive answer



Hello,

thanks for the answers to my question. In fact I had allready 
read the documentation, but measuring with the rpcc failed to find a 
instruction sequence that needed less than 1 cycle/instruction.

But it is possible indeed as:

addq, addq, mult, addt, addq, addq can be execute in 2 or 3 cycles

if there are no register dependencys. 
Note: this is measured not deducted from the manual.

Anyway I think I have to make a few more passes on the documentation
to find out how a given sequence is scheduled exactly.


Matti Aarnio <matti.aarnio@sonera.fi> wrote:

> No, in fact they take more -- FP multiplier has pipeline
>         latency of 8 cycles.  FP-add pipeline takes also 8 cycles.
> (archive digging --- dec21164pchwm.pdf found ...)

My documentation said FP mul/add takes 4 cycles latency

Oscar Levi <elf@buici.com> wrote:

> I'm not sure you would see any given fragment execute in less than 4
> cycles.  Instead, you might see the aggregate performance of a
> sequence of instructions be less than the sequential fetch time.

That's exactly what I was searching for. A explicit written down sequence
that actually execute in less cycles than instructions.

Robert Harley <Robert.Harley@inria.fr> wrote:

> In answer to the question: yes the 21164 can issue up to four
> instructions in one cycle under the right conditions.  The example
> given looks to me like it ought to go through in one as long as it
> is correctly aligned, but check the document mentioned in the "ALPHA
> instruction timing" thread a few days ago for a definitive answer.

Yes, I read the documentation I that that is the kind of sequence that
should run at maximum speed, because all instructions can use
a different execution unit and no register resource conflicts are
there.

BUT I measured the time using the rpcc and it says 4 cycles.
BUT what is rpcc exactly measuring?

"B. James Phillippe" <bryan@terran.org> wrote:

> We all hate repeat questions about documented information, probably as much
> as we hate getting RTFM responses. ;)  Would the other list members object
> to an automated monthly posting containing URL's to FAQ's and other sources
> of Linux/AXP documentation (our fantastic websites, etc)?  I would
> volunteer to do it if nobody else does (assuming the overhwelming majority
> of list members like the idea).  The message would not need to be large to
> be effective; even a paragraph or so would suffice, I think.

Good idea. I voluntieer to contribute a little assembly code
example. ;-)

Greetings
Thorsten
-- 
"They that can give up essential liberty to obtain a little temporary
saftey deserve neither liberty not saftey."
-- Benjamin Franklin, 1759



[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index] []