[cairo] Pixman - refactoring of fbFetchTransformed - performance issue

[Bertram Felgenhauer]

Antoine Azar wrote:
> I'm very surprised at this 30% number. The refactoring shouldn't come 
> anywhere close to that number.

I agree, and I was surprised as well. However, the result is
reproducible and the numbers are stable.

It appears that the compiler (gcc) does a worse job optimizing the
refactored code. One difference is that it needs to set up a larger
stack frame for fetchFromRegion() than it had to for fetch().

On the other hand, the additional indirect call is as cheap as possible;
the whole code of fbFetchFromNoRegion is

00005320 <fbFetchFromNoRegion>:
    5320:	8b 4c 24 14		mov    0x14(%esp),%ecx
    5324:	ff e1                	jmp    *%ecx

and the jump should be predicted nicely.

I haven't looked at the differences in detail.

> I applied Bertram's patch for the perf test suite and ran just those 
> mag/min tests before and after the refactoring, and I'm seeing 1-2% 
> performance differences. You can see the results of both runs and the 
> speedup/slowdown factor here: http://www.antoineazar.com/refactor.htm

What processor and which compiler (and options) are you using?


As an aside, it's odd that the difference between *_source and *_over
is so large in your tests, and that *_source turns out to be slower.
I wonder why.

My numbers are different, for example:

(before)
[  9]    image-rgba  paint_image_rgba_mag_source-256   12614763    6.919    7.006  0.81%   8
[ 10]    image-rgba   paint_image_rgba_min_over-256   14194059    7.785    7.788  0.09%   4

(after)
[  9]    image-rgba  paint_image_rgba_mag_source-256   16889333    9.264    9.267  0.02%   5
[ 10]    image-rgba   paint_image_rgba_min_over-256   18351320   10.066   10.086  0.51%   5

with hardly any difference between _source and _over.

> Could a third person run the perf before and after the refactoring to 
> check the numbers?

Yes, that would help - the more numbers the better.

The relevant commits are in the pixman repository,
   e95638c629334151e27633cc1c476ea582d766ec
   (before the refactoring)
and
   8d79c48126398aa7b31e9bb9e25af9d231075604
   (after)

Of particular interest are the *_image_rgba_min_* and
*_image_rgba_mag_* tests in the performance testsuite (now in
cairo's git repository).

Bertram