[cairo] Added mmx fast path for fbComposeCopyAreammx

[Frédéric Plourde]

Soeren Sandmann a écrit :
> Vladimir Vukicevic <vladimir at pobox.com> writes:
>
>   
>> On Mar 10, 2008, at 12:21 PM, Frederic Plourde wrote:
>>     
>>> Here's a speedup chart that supports this optimization :
>>> http://pastebin.mozilla.org/362713
>>> A mean perf gain of 1.5X was observed for the image-rgb cases  
>>> instead of 0.99 as
>>> seen before this optim.
>>>       
>> There's something weird here, everything shows a speedup except:
>>
>> before patch:
>> image-rgb subimage_copy-512    0.303ms  (nommx) -> 0.321ms (mmx)  --  
>> 94% speed
>>
>> after patch:
>> image-rgb subimage_copy-512    0.302ms  (nommx) -> 0.347ms (mmx)  --  
>> 87% speed
>>     
>
> Notice that number of iterations is 7. That is just not going to give
> meaningful numbers.
>
> Here is what I have been doing for testing pixman performance:
>
> (1) Apply this patch:
>
>         http://www.daimi.au.dk/~sandmann/stats2.patch
>
>     This stops cairo from computing statistics in between
>     iterations.
>
> (2) Run something like 
>
>         env CAIRO_TEST_TARGET=image cairo-perf -i 100000 subimage_copy > before.perf
>
> (3) Apply pixman changes, and run make install.
>
> (4) Then run
>
>         env CAIRO_TEST_TARGET=image cairo-perf -i 100000 subimage_copy > after.perf
>
> (5) Then do this
>
>         cairo-perf-diff before.perf after.perf
>
> For me, with this procedure, this patch makes no significant
> difference. I did verify that it does get hit.
>
> This is not too surprising though, as without the patch we just
> hitmemcpy(). In both cases, we are likely to simply be memory
> bandwidth limited.
>
>
> Soren
>   

Thanks for the tip Soeren.
However, I've run 500 iterations on the image targets with subimage_copy 
perf tests... got very decent STD levels, and still, I got significant 
speedups !..
see the speedup chart : http://pastebin.mozilla.org/366980

(weird that we get different results... you're running it under linux, 
right ? That could explain the difference.. I'll run it under linux as 
well and compare.)

Plus I intend to take a closer look to larger-sized images where the 
perf gains seem to drop ! (This might catch Vald's interest ;-)
get back to you later ;-)
-fred-