Fix for Ovals/RRects on N4 missing interior rectangles.

Fix for Ovals/RRects on N4 missing interior rectangles.

Clamp square of gradiant length to non-zero value so it can be passed to
inversesqrt().
Also adds some fixes to rrect setup to avoid crossing zero.

Skia issue #1372

R=bsalomon@google.com, robertphillips@google.com

Committed: https://code.google.com/p/skia/source/detail?r=9864

[jvanverth1, 2013-07-01 18:32:17]

No real performance difference on MacOS with this change:

rrects_3_stroke_4 GPU           13.11      13.17      -0.06    -0.5% 
          ovals_1 GPU            1.00       1.00      +0.00    +0.0% 
          ovals_3 GPU            1.00       1.00      +0.00    +0.0% 
 ovals_3_stroke_4 GPU           13.00      13.00      +0.00    +0.0% 
         rrects_1 GPU            1.01       1.01      +0.00    +0.0% 
rrects_1_stroke_4 GPU           19.00      19.00      +0.00    +0.0% 
         rrects_3 GPU            1.00       1.00      +0.00    +0.0% 
 ovals_1_stroke_4 GPU           18.79      18.69      +0.10    +0.5%

[robertphillips, 2013-07-01 18:35:41]

lgtm

https://codereview.chromium.org/18375003/diff/1/src/gpu/GrOvalRenderer.cpp
File src/gpu/GrOvalRenderer.cpp (right):

https://codereview.chromium.org/18375003/diff/1/src/gpu/GrOvalRenderer.cpp#ne...
src/gpu/GrOvalRenderer.cpp:224: builder->fsCodeAppend("\tfloat grad_dot =
dot(grad, grad);\n");
comment here as to why we're doing this?

[bsalomon, 2013-07-01 18:36:02]

Should we try measuring the perf on a mobile GPU where perf may be more
sensitive to conditionals? Perhaps on SGX?

https://codereview.chromium.org/18375003/diff/1/src/gpu/GrOvalRenderer.cpp
File src/gpu/GrOvalRenderer.cpp (right):

https://codereview.chromium.org/18375003/diff/1/src/gpu/GrOvalRenderer.cpp#ne...
src/gpu/GrOvalRenderer.cpp:225: builder->fsCodeAppend("\tgrad_dot =
max(grad_dot, 1.0e-4);\n");
Could use a comment here about why this number was selected.

[jvanverth1, 2013-07-01 18:42:47]

On 2013/07/01 18:36:02, bsalomon wrote:
> Should we try measuring the perf on a mobile GPU where perf may be more
> sensitive to conditionals? Perhaps on SGX?

I think max() will take the same amount of time whether it takes the branch or
not, but I'll check.

> https://codereview.chromium.org/18375003/diff/1/src/gpu/GrOvalRenderer.cpp
> File src/gpu/GrOvalRenderer.cpp (right):
> 
>
https://codereview.chromium.org/18375003/diff/1/src/gpu/GrOvalRenderer.cpp#ne...
> src/gpu/GrOvalRenderer.cpp:225: builder->fsCodeAppend("\tgrad_dot =
> max(grad_dot, 1.0e-4);\n");
> Could use a comment here about why this number was selected.

Will do.

[jvanverth1, 2013-07-01 22:25:38]

Here are the results for ovals, running on the nexus-s: 

         ovals_1 GPU          121.57     129.34      -7.77    -6.4% 
         ovals_3 GPU           19.90      20.28      -0.38    -1.9% 
ovals_1_stroke_4 GPU          345.00     349.45      -4.45    -1.3% 
ovals_3_stroke_4 GPU          114.23     115.13      -0.90    -0.8% 


On 2013/07/01 18:42:47, JimVV wrote:
> On 2013/07/01 18:36:02, bsalomon wrote:
> > Should we try measuring the perf on a mobile GPU where perf may be more
> > sensitive to conditionals? Perhaps on SGX?
> 
> I think max() will take the same amount of time whether it takes the branch or
> not, but I'll check.
> 
> > https://codereview.chromium.org/18375003/diff/1/src/gpu/GrOvalRenderer.cpp
> > File src/gpu/GrOvalRenderer.cpp (right):
> > 
> >
>
https://codereview.chromium.org/18375003/diff/1/src/gpu/GrOvalRenderer.cpp#ne...
> > src/gpu/GrOvalRenderer.cpp:225: builder->fsCodeAppend("\tgrad_dot =
> > max(grad_dot, 1.0e-4);\n");
> > Could use a comment here about why this number was selected.
> 
> Will do.

[bsalomon, 2013-07-02 13:09:33]

On 2013/07/01 22:25:38, JimVV wrote:
> Here are the results for ovals, running on the nexus-s: 
> 
>          ovals_1 GPU          121.57     129.34      -7.77    -6.4% 
>          ovals_3 GPU           19.90      20.28      -0.38    -1.9% 
> ovals_1_stroke_4 GPU          345.00     349.45      -4.45    -1.3% 
> ovals_3_stroke_4 GPU          114.23     115.13      -0.90    -0.8% 
> 
> 
> On 2013/07/01 18:42:47, JimVV wrote:
> > On 2013/07/01 18:36:02, bsalomon wrote:
> > > Should we try measuring the perf on a mobile GPU where perf may be more
> > > sensitive to conditionals? Perhaps on SGX?
> > 
> > I think max() will take the same amount of time whether it takes the branch
or
> > not, but I'll check.
> > 
> > > https://codereview.chromium.org/18375003/diff/1/src/gpu/GrOvalRenderer.cpp
> > > File src/gpu/GrOvalRenderer.cpp (right):
> > > 
> > >
> >
>
https://codereview.chromium.org/18375003/diff/1/src/gpu/GrOvalRenderer.cpp#ne...
> > > src/gpu/GrOvalRenderer.cpp:225: builder->fsCodeAppend("\tgrad_dot =
> > > max(grad_dot, 1.0e-4);\n");
> > > Could use a comment here about why this number was selected.
> > 
> > Will do.
lgtm

[jvanverth1, 2013-07-02 14:46:09]

Committed patchset #1 manually as r9864 (presubmit successful).