[OpenJDK 2D-Dev] RFR JDK-8184429: Path clipper added in Marlin2D & MarlinFX 0.8.0

[Laurent Bourgès]

Phil & Kevin,
The proposed plan looks good.

I will first work on refactoring my ShapeClipTest and then focus on having
a Java2d patch ready.

I absolutely want that this new performance improvement will ship in Java
10.

PS: Another important performance change concerns tuning Java2D tile sizes.
This can be easily done: set tile log2 settings to 7 x 6 to let Marlin use
128x64 tiles: up to 60% gain for large shapes rendered on VolatileImage
(linux 64 + i7 + nvidia)

I will create new bugs to track these RFE...

Bye,
Laurent

Le 8 nov. 2017 23:44, "Phil Race" <philip.race at oracle.com> a écrit :

I think they should be separate webrevs sent to the separate lists and you
should start with 2D
as I can then run the JDK regression tests on it. I know you can
theoretically run the open regression
tests too (are you ?) but there are some random scattered closed regression
tests that so far
as I can see can be open sourced .. that I can run but you can't .. I'll at
least run the
automated ones. I wouldn't call them anything very focused on testing
rasterization but
I can at least check off that concern ..

And yes, I'll make time to review it.

-phil.


On 11/08/2017 01:55 PM, Laurent Bourgès wrote:

Kevin & Phil,

Some news on that issue:
I successfully managed to finish the Path clipping support in Marlin 0.8.2
(release last week):
https://github.com/bourgesl/marlin-renderer/releases/tag/v0.8.2

I fixed few remaining bugs in either Stroker (1) and in PathClipFilter (2)
to have proper & tested clipping in Marlin renderer (2D). It now works
perfectly with either NZ or EO winding rules.

To ensure detecting any artefact between Clipping Off vs On, I implemented
a 'basher' test (as recommended by Jim) that renderers 10 000 random
polygons (5 -> 9 -> 50 line segments or mixed with line / quads / cubics) (
whose point coordinates are in [-50 to 150] ) to a 100x100 buffered image
with or without clipping enabled (using a system property at runtime). Of
course, all output pixels are compared and any pixel difference is
considered as a failure.

The new ShapeClipTests tests all stroke combinations (cap / join / with or
without dashes / closed or not / EO or NZ rule) and also fills (closed or
not / EO or NZ rule) => 170 tests run OK

I need some time to synchronize MarlinFX and then with either OpenJDK
forrest (new) or OpenJFX10.
If you want the new automated test (long run ~ 20 minutes), I need some
time to refactor it as it uses some code from my MapBench tool and have a
standalone test class.

Will you have time to review such (medium) changes in Marlin2D (Phil ?) and
/ or MarlinFX (Kevin ?) before the deadline (dec 14th) ?
I said 'medium' as the code is quite simple to read but the new CG
algorithms to ignore / discard useless path elements are cool but not
obvious.

Please tell me if you have time and if you prefer a combined (JDK / JFX)
webrev or start with 2D or JFX.

Cheers,
Laurent


2017-09-07 8:52 GMT+02:00 Laurent Bourgès <bourges.laurent at gmail.com>:

> Hi Kevin,
>
> Ok I propose to withdraw or postpone this review after JavaOne where we
> will be able to discuss in a face to face meeting about Marlin & MarlinFX
> changes for JDK10.
>
> I hope the 2d / jfx groups have other Graphics Guru to help, as good as
> Jim Graham.
>
> Cheers,
> Laurent
>
> Le 6 sept. 2017 16:23, "Kevin Rushforth" <kevin.rushforth at oracle.com> a
> écrit :
>
>> Hi Laurent,
>>
>> Some combination of Phil, Sergey, and I will take a look at this when we
>> can. Perhaps there might be others on these two lists who could lend a
>> helping hand?
>>
>> -- Kevin
>
>