In convex path renderer, translate polygon to origin for fanPt computation

[bsalomon, 2012-02-23 18:36:57]

This is an alternate CL that fixes the same issue as:
http://codereview.appspot.com/5672069/ (Skia Issue 497)

I did some timings of both. There is no measurable difference on my Linux
machine but I'm a little hesitant to promote to double on all architectures.

An interesting note: The issue doesn't reproduce on Windows. We use more
aggressive fp optimization flags there. I believe that intermediate values are
kept in higher precision registers rather than truncated. Setting -ffast-math on
Linux makes the issue much better with the original code.

[reed1, 2012-02-23 18:45:20]

http://codereview.appspot.com/5698051/diff/2001/src/gpu/GrAAConvexPathRendere...
File src/gpu/GrAAConvexPathRenderer.cpp (right):

http://codereview.appspot.com/5698051/diff/2001/src/gpu/GrAAConvexPathRendere...
src/gpu/GrAAConvexPathRenderer.cpp:76: pi = pj;
do we need the '% count' here?

[bsalomon, 2012-02-23 18:50:27]

http://codereview.appspot.com/5698051/diff/2001/src/gpu/GrAAConvexPathRendere...
File src/gpu/GrAAConvexPathRenderer.cpp (right):

http://codereview.appspot.com/5698051/diff/2001/src/gpu/GrAAConvexPathRendere...
src/gpu/GrAAConvexPathRenderer.cpp:76: pi = pj;
On 2012/02/23 18:45:21, reed1 wrote:
> do we need the '% count' here?

Good point, it's no longer necessary. Removed in the latest patch.

[reed1, 2012-02-23 19:09:24]

lgtm

[bsalomon, 2012-02-23 19:21:04]

Committed as r3241. guanqun, thanks for finding this and suggesting a fix.
Please let me know if you think the bug isn't adequately fixed.

[guanqun, 2012-02-24 01:37:34]

I understand your concerns to use "double" on all architectures. Thanks for your
new fixup. It solves the problem.