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

On Aug 5, 5:29 am, Martin Brown <|||newspam...@xxxxxxxxxxxxxxxxxx>
wrote:
aruzinsky wrote:
I forgot prefetchN = 8.

On Aug 4, 11:42 am, aruzinsky <aruzin...@xxxxxxxxxxxxxxxxxxxx> wrote:
On Aug 4, 12:15 am, Hendrik van der Heijden <h...@xxxxxx> wrote:

aruzinsky schrieb:
Thank you for your input.  I would feel more confident if you tested
my code, though.

Exact behaviour will depend on which CPU architecture it runs on.

Then post a pointer to your code, sources preferred.
Hendrik
Simple enough to cut and paste here, but this is an untested excerpt
that assumes the arrays are a multiple of 16 floats aligned on 16 byte
boundaries.

Only really worth looking at the innermost loop. It hits [edx +
64*prefetchN] with prefetch hints every time around. I confess some
surprise that it makes so much difference.

[snip]





                align 16;
$L2:
                movaps  xmm0, [edx]
                movaps  xmm1, [edx+16]
                movaps  xmm2, [edx+32]
                movaps  xmm3, [edx+48]

                PREFETCHNTA [edx+ecx]

                movntpd  [eax], xmm0
                movntpd  [eax+16], xmm1
                movntpd  [eax+32], xmm2
                movntpd  [eax+48], xmm3

                add     edx, 64
                add     eax, 64

                dec esi
                jnz     $L2

At first glance there is a possible [AGI] partial stall on the first
instruction in the loop depending on how good the branch
prediction/speculative execution is. Try moving add edx, 64 up a bit.

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.

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.


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....

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

- Show quoted text -

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

Incidentally, everyone, here is a SSE math library without
prefetches:

http://www.applied-mathematics.net/miniSSEL1BLAS/








.