Re: Share Your Experience with 3DNow, SSE, SSE2 etc.

On Aug 7, 7:17 am, Martin Brown <|||newspam...@xxxxxxxxxxxxxxxxxx>
wrote:
aruzinsky wrote:
On Aug 5, 5:29 am, Martin Brown <|||newspam...@xxxxxxxxxxxxxxxxxx>
wrote:

Did you establish experimentally that a loop unrolled by 4 was optimum?

I think so. Maybe, I copied it from a similar function that I
optimized.

Worth checking. Cache behaviour can be rather quirky at times.







Tempting to suggest the usual peephole optimisations, and also to try
out the shortest loop to see how well or badly that performs. eg.
     $L2:
                    movaps xmm0, [esi]
                   movntpd [ebx+esi], xmm0
                    add esi, 16
                    dec ecx                  ; or loop $L2
                    jnz $L2
Registers used all different ebx contains old(eax-edx) etc.
And then try loop unrolling.

But rather than optimise by trial and error, blindfolded why not enable
the performance monitoring counters and do it properly by monitoring
cache misses and stalls.

But, I still have to use trial and error to find the best code
arrangement for prefetches because prefretches should be interleaved
with computations.

If you are serious about coding for multiple CPUs with near optimality
then the parameterised generated codelets approach used by FFTW and
others is probably the way to go. See Bugbears post for details.



The Ring0 driver ia32.sys is at the University of Texas site (playing up
at the moment) but I think Google cache still has a copy - including a
new class library around it.

Try at your own peril but it can be very useful for tuning critical code.

http://216.239.59.104/search?q=cache:H9fYu1GGy0EJ:iss.ices.utexas.edu....

Thank you.  I take it that you no longer have easy access to a C++
compiler to do your own experiments?

More lack of time than anything else.
I would only optimise at this sort of low level as a last resort. YMMV

Regards,
Martin Brown
** Posted fromhttp://www.teranews.com**- Hide quoted text -

- Show quoted text -

I checked and the loop unroll was unnecessary. I still get a 6% speed
increase with prefetchnta.

I can get a 20% speed increase for conjugate gradients which my
software, SAR Image Processor, often uses, but advertising SIMD would
mostly serve as a sales gimmick.

To do better, I would have to use 2D float arrays with all rows
beginning on 16 byte boundaries. Rewriting code for such arrays would
be a major headache because my current matrix functions assume that
rows are contiguous in RAM.
.

Relevant Pages

  • Re: C# coding guidelines: use "this." or not when referring to member fields/properties within the
    ... Alphabetic for strings isn't quite so ... One example of where people go wrong is when they want to optimise loop ... implementation so that each iteration takes 10% less time will only ...
    (microsoft.public.dotnet.languages.csharp)
  • Re: hot path optimizations in uma_zalloc() & uma_zfree()
    ... > I suppose the reason of first gain lies in increasing of cpu cache hits. ... > separate buckets. ... I ran ministat against your tests with 1000 sockets loop and there isn't a lot ...
    (freebsd-hackers)
  • Re: hot path optimizations in uma_zalloc() & uma_zfree()
    ... > I suppose the reason of first gain lies in increasing of cpu cache hits. ... > separate buckets. ... I ran ministat against your tests with 1000 sockets loop and there isn't a lot ...
    (freebsd-hackers)
  • [PATCH][RFC] fast file mapping for loop
    ... are done once they hit page cache. ... loop without making it even slower than it currently is. ... * Add bio to back of pending list and wakeup thread ... * Find extent mapping this lo device block to the file block on the real ...
    (Linux-Kernel)
  • Cache size restrictions obsolete for unrolling?
    ... into the cache (i.e. the code of the loops was slightly smaller than ... execution time using a cycle-true simulator. ... unrolling factor stepwise resulting in the unrolled loop that exceeded ... I expected to get a performance decrease, i.e. the stronger the loop ...
    (comp.compilers)