Re: [11] JDK-8204621: Upgrade MarlinFX to 0.9.2

Fri, 29 Jun 2018 11:27:35 -0700

One more thing about the test. All of the OpenJFX unit tests should have GPL v2 + Classpath Exception (this differs from the JDK). 

-- Kevin


On 6/29/2018 10:23 AM, Kevin Rushforth wrote:
I'll plan to review the code today if possible. This will need one more reviewer, so maybe Phil can also review it, since he reviewed the Java2D patch? 

As for my comments on the test:
Finally I think this test should be manually run only if Marlin renderer is modified. 
How to do that ? use @Ignore or specific tags ?
As a slight variation of this: How about running a small number (say, 200 or 250) by default, but adding a flag to run more? Alternatively, you could use a flag to enable it, but since you would need to do something extra (provide a flag or modify the test) to run it, we might as well get at least some testing all the time. Unless you really think there is no value in doing this.
I deliberately set all these Marlin clip (runtime + always subdivider) / curve quality settings (quads / cubics thresholds) to be sure of the concrete Marlin setup as quality thresholds are sensitive to such values.
As a best practice, tests should generally be run using the same settings as are used in production. Other than to verify how it behaves when you change these settings, I don't see the value in testing the system running in a mode that no application will ever see. I may be missing some point here. 

-- Kevin


On 6/25/2018 9:01 AM, Laurent Bourgès wrote:

Kevin,

Here are my comments below:
2018-06-16 1:47 GMT+02:00 Kevin Rushforth <kevin.rushfo...@oracle.com <mailto:kevin.rushfo...@oracle.com>>: 

    I tested this on all three platforms and the updated rasterizer
    looks good.

    I spot checked the code changes, but didn't get time to do a
    complete review yet. I was mostly looking for diffs between the
    Java2D version which was already reviewed, and this one.

    I do have a couple comments on the new ClipShapeTest (which looks
    like a nice accuracy test, btw).

    1. The test runs for way too long (about 20x too long) to include
    in our normal test runs. By default the entire class file (all
    three tests) needs to take < 5 minutes and 2 minutes would be
    better. I measured the time on 4 machines that I have and found
    that if you cut the number of iterations down from 5000 to 250 it
    will be just about the right run time. Then you can set the
    timeout to 120 seconds (the slowest test on the slowest of my
    machines took about 48 seconds, so a 2 minute timeout should be
    plenty).
I agree this test is very long but it is the only mean I found to test all possible stroke combinations and test enough shapes (5000) to detect bugs. I wondered if using mask directly (via ShapeUtils.getMaskData()) would become faster but it will never run below the 2 minutes threshold in total.
Finally I think this test should be manually run only if Marlin renderer is modified. 
How to do that ? use @Ignore or specific tags ?


    2. Can you explain the reason for setting the following?

     206         // disable static clipping setting:
     207         System.setProperty("prism.marlin.clip", "false");
     208
    System.setProperty("prism.marlin.clip.runtime.enable", "true");
     209
     210         // enable subdivider:
     211 System.setProperty("prism.marlin.clip.subdivider",
    "true");
     212
     213         // disable min length check: always subdivide curves
    at clip edges
     214 System.setProperty("prism.marlin.clip.subdivider.minLength",
    "-1");
     215
     216         // If any curve, increase curve accuracy:
     217         // curve length max error:
     218 System.setProperty("prism.marlin.curve_len_err", "1e-4");
     219
     220         // cubic min/max error:
     221         System.setProperty("prism.marlin.cubic_dec_d2", "1e-3"); 
     222         System.setProperty("prism.marlin.cubic_inc_d1",
    "1e-4"); // or disabled ~ 1e-6
     223
     224         // quad max error:
     225         System.setProperty("prism.marlin.quad_dec_d2", "5e-4");

    It seems better to test with the default parameters (i.e., it
    makes a better regression test that way).
I deliberately set all these Marlin clip (runtime + always subdivider) / curve quality settings (quads / cubics thresholds) to be sure of the concrete Marlin setup as quality thresholds are sensitive to such values.
The ClipShapeTest is dedicated to test the clipper (+ subdivider) part of the Marlin renderer. 



    3. Related to that, I think you should eliminate the following (I
    don't recommend running functional tests with this set), although
    since you don't do any animation, it probably doesn't matter.

     227 System.setProperty("javafx.animation.fullspeed",
    "true"); // full speed


I will remove it and see if the overall test is not slower.
Is Platform.runLater() impacted by such setting (latency of FX thread -> Prism rendering thread ?) ? 

Laurent



    On 6/8/2018 7:28 AM, Laurent Bourgès wrote:

        Hi,

        Please review this large patch to upgrade MarlinFX to 0.9.2 in
        OpenJFX11:
        JBS: https://bugs.openjdk.java.net/browse/JDK-8198885
        <https://bugs.openjdk.java.net/browse/JDK-8198885>
        webrev:
http://cr.openjdk.java.net/~lbourges/marlinFX/marlinFX-092.0/
<http://cr.openjdk.java.net/%7Elbourges/marlinFX/marlinFX-092.0/>
<http://cr.openjdk.java.net/%7Elbourges/marlinFX/marlinFX-092.0/
<http://cr.openjdk.java.net/%7Elbourges/marlinFX/marlinFX-092.0/>>
        PR: https://github.com/javafxports/openjdk-jfx/pull/96
<https://github.com/javafxports/openjdk-jfx/pull/96> (CI OK)

        This patch is almost identical to Marlin(2D) patch, see:
        https://bugs.openjdk.java.net/browse/JDK-8198885
        <https://bugs.openjdk.java.net/browse/JDK-8198885>

        I added the ClipShapeTest (ported to jfx) that compares shape
        clipping (within threshold) and it works (within large timeouts): 
        gradle -PFULL_TEST=true :system:test --tests
        test.com.sun.marlin.ClipShapeTest

        Regards,
        Laurent

Reply via email to