[OpenJDK 2D-Dev] JDK-8220154 Improve java2d rendering performance on macOS by using Metal framework

Ajit Ghaisas ajit.ghaisas at oracle.com
Mon Jul 8 12:25:32 UTC 2019


Hi,

   I have pushed a minor correction to the metal-prototype-branch of jdk-sandbox (http://hg.openjdk.java.net/jdk/sandbox <http://hg.openjdk.java.net/jdk/sandbox>)
   
   Now, we can build and launch J2DDemo, J2DBenchmark and SwingSet2 with Metal pipeline.
   Of course, rendering performance is slow and still a lot of functionality needs to be implemented.

Regards,
Ajit


> On 20-Jun-2019, at 3:50 AM, Alexey Ushakov <alexey.ushakov at jetbrains.com> wrote:
> 
> Hi Ajit,
> 
>>     Thanks for suggesting how to run basic perf test.
>>     I did run it with and without Metal. The test run with Metal takes ~2X time to run as compared with OpenGL.
>>     What is your observation?
> 
> Yes, in some cases we have similar performance drop at the moment (testFlatBubbles, testFlatOvalRotBubbles, testWiredBubbles, testFlatQuad, testWiredQuad)
> Here is the results from Mac mini 2018, macOS 10.14.4
> 
> metalEnabled=0
> testFlatBubbles 70
> testFlatBoxBubbles 78
> testImgBubbles 73
> testFlatBoxRotBubbles 78
> testFlatOvalRotBubbles 74
> testLinGradOvalRotBubbles 46
> testWiredBubbles 54
> testWiredBoxBubbles 82
> testLines 84
> testFlatQuad 62
> testWiredQuad 58
> 
> metalEnabled=1
> testFlatBubbles 31
> testFlatBoxBubbles 91
> testImgBubbles 88
> testFlatBoxRotBubbles 92
> testFlatOvalRotBubbles 36
> testLinGradOvalRotBubbles 37
> testWiredBubbles 17
> testWiredBoxBubbles -
> testLines 89
> testFlatQuad 21
> testWiredQuad 19
> 
>>     Although, the java side code changes are good, I see an issue with native Metal renderer implementation. 
>>     For rendering primitives, I see that the metal commands are being encoded for each drawXXX call in MTLRenderer.m
>>  
>>     Why is MetalContext.createRenderEncoderForDest is being created for each draw call? I think, it is an overhead.
>>     Have you considered managing a VertexBuffer and encoding ’N’ set of draw calls in a render pass?
> 
> Yes, I thought about such optimizations. They definetely should help with this performance drop.
> 
> Best Regards,
> Alexey  
> 
> 
>> On 11 Jun 2019, at 12:01, Ajit Ghaisas <ajit.ghaisas at oracle.com <mailto:ajit.ghaisas at oracle.com>> wrote:
>> 
>> Hi Alexey,
>> 
>>     Thanks for suggesting how to run basic perf test.
>>     I did run it with and without Metal. The test run with Metal takes ~2X time to run as compared with OpenGL.
>>     What is your observation?
>> 
>>     Although, the java side code changes are good, I see an issue with native Metal renderer implementation. 
>>     For rendering primitives, I see that the metal commands are being encoded for each drawXXX call in MTLRenderer.m
>>  
>>     Why is MetalContext.createRenderEncoderForDest is being created for each draw call? I think, it is an overhead.
>>     Have you considered managing a VertexBuffer and encoding ’N’ set of draw calls in a render pass?
>> 
>> Regards,
>> Ajit
>> 
>> 
>>> On 17-May-2019, at 12:20 PM, Jayathirth Rao <jayathirth.d.v at oracle.com <mailto:jayathirth.d.v at oracle.com>> wrote:
>>> 
>>> Hi Alexey,
>>> 
>>> Thanks for the clarifications.
>>> Build failures we are getting are mainly because of tighter compiler restrictions, I have attached error log for the same.
>>> Basically we have 2 log errors and one conditional error.
>>> 
>>> Yes in the latest build there is refactoring of LWComponentPeer.java but I have taken care of it and build error was not because of that.
>>> 
>>> And thanks for sharing test and Graphics2D supported method details.
>>> 
>>> Regards,
>>> Jay
>>> <Build_failure.txt>
>>> 
>>> 
>>>> On 16-May-2019, at 10:39 PM, Alexey Ushakov <alexey.ushakov at jetbrains.com <mailto:alexey.ushakov at jetbrains.com>> wrote:
>>>> 
>>>> Hi Jay,
>>>> 
>>>>>> I can see that now you have added the MTLTexturePool.m file. It only partially solves the build failures.
>>>>>> Jay & myself tried building it separately and we do still see the build failures.
>>>> 
>>>> It’s strange. There wasn’t any copy/paste errors. The only reason that I can imagine that some chunks of the patch were rejected (as I actually experienced recently because of some refactoring of LWComponentPeer.java happened couple of days ago).
>>>> Anyway, I regenerated webrev once more (http://cr.openjdk.java.net/~avu/JDK-8220154/webrev.01 <http://cr.openjdk.java.net/~avu/JDK-8220154/webrev.01>) and verified it with the openjdk workspace hg.openjdk.java.net/jdk/client <http://hg.openjdk.java.net/jdk/client> . 
>>>> 
>>>> 
>>>>> So with current JB changes what parts of primitive and image drawing are expected to work?
>>>>> Please provide your inputs.
>>>> 
>>>> I’ve included junit4 performance test that you can try to see what is currently supported.
>>>> Here is some steps to run the test provided with the webrev
>>>> 
>>>> #cd openjdk_ws 
>>>> #mkdir -p test/jdk/jbu/testdata/perf/metal/
>>>> #cp webrev/raw_files/new/test/jdk/jbu/testdata/perf/metal/duke.png test/jdk/jbu/testdata/perf/metal/
>>>> 
>>>> #./build/macosx-x86_64-server-release/images/jdk-bundle/jdk-13.jdk/Contents/Home/bin/javac  -cp .:/Users/avu/tools/junit-4.10.jar -d . test/jdk/jbu/perf/metal/MetalPerfTest.java
>>>> 
>>>> #./build/macosx-x86_64-server-release/images/jdk-bundle/jdk-13.jdk/Contents/Home/bin/java  -cp .:/Users/avu/tools/junit-4.10.jar -Dtestdata=/Users/avu/ws/openjdk/test/jdk/jbu/testdata -Djb.java2d.metal=true org.junit.runner.JUnitCore  perf.metal.MetalPerfTest
>>>> 
>>>> Here is the list of supported methods from Graphics2D: 
>>>> setColor
>>>> fillOval
>>>> translate
>>>> setTransform
>>>> rotate
>>>> setPaint(new LinearGradientPaint())
>>>> drawImage
>>>> fillRect
>>>> draw/fill (Shape)
>>>> 
>>>> Best Regards,
>>>> Alexey 
>>>> 
>>>>> On 16 May 2019, at 15:08, Jayathirth Rao <jayathirth.d.v at oracle.com <mailto:jayathirth.d.v at oracle.com>> wrote:
>>>>> 
>>>>> Hi Alexey,
>>>>> 
>>>>> We are trying to test basic calls of Graphics2D as mentioned in https://docs.oracle.com/javase/tutorial/2d/TOC.html <https://docs.oracle.com/javase/tutorial/2d/TOC.html> 
>>>>> 
>>>>> I see that Graphics2D.drawImage() with BufferedImage as input works. Also I tried other operations like fillRect() and fillOval() and they work. I am testing with simple JFrame and adding draw calls inside overriden paint() method of a panel.
>>>>> 
>>>>> So with current JB changes what parts of primitive and image drawing are expected to work?
>>>>> Please provide your inputs.
>>>>> 
>>>>> Thanks,
>>>>> Jay
>>>>> 
>>>>>> On 16-May-2019, at 12:47 PM, Ajit Ghaisas <ajit.ghaisas at oracle.com <mailto:ajit.ghaisas at oracle.com>> wrote:
>>>>>> 
>>>>>> Thanks Alexey.
>>>>>> 
>>>>>> I can see that now you have added the MTLTexturePool.m file. It only partially solves the build failures.
>>>>>> Jay & myself tried building it separately and we do still see the build failures.
>>>>>> 
>>>>>> Some statements need bracketing & some copy-paste errors need correction.
>>>>>> We could fix these errors and have got a build. We will continue our test experiments with it.
>>>>>> 
>>>>>> Meanwhile, you may want to update the patch to fix the build errors. 
>>>>>> 
>>>>>> Regards,
>>>>>> Ajit
>>>>>> 
>>>>>> 
>>>>>>> On 08-May-2019, at 8:56 PM, Alexey Ushakov <alexey.ushakov at jetbrains.com <mailto:alexey.ushakov at jetbrains.com>> wrote:
>>>>>>> 
>>>>>>> FYI, I’ve rebased our work on top of the latest state of http://hg.openjdk.java.net/jdk/jdk <http://hg.openjdk.java.net/jdk/jdk> 
>>>>>>> (http://cr.openjdk.java.net/~avu/JDK-8220154/webrev.01 <http://cr.openjdk.java.net/~avu/JDK-8220154/webrev.01>)
>>>>>>> 
>>>>>>> Best Regards,
>>>>>>> Alexey
>>>>>>> 
>>>>>>>> On 8 May 2019, at 16:19, Alexey Ushakov <Alexey.Ushakov at jetbrains.com <mailto:Alexey.Ushakov at jetbrains.com>> wrote:
>>>>>>>> 
>>>>>>>> Thanks for catching it, Ajit!
>>>>>>>> 
>>>>>>>> Looks like it was a problem with webrev script applied to multiple git commits. I’ve updated the webrev.
>>>>>>>> Also, we didn’t rebase yet on the latest state of http://hg.openjdk.java.net/jdk/jdk <http://hg.openjdk.java.net/jdk/jdk> (this work is in progress).
>>>>>>>> I’ll let you know when the rebase is ready.
>>>>>>>> 
>>>>>>>> Best Regards,
>>>>>>>> Alexey
>>>>>>>> 
>>>>>>>>> On 7 May 2019, at 21:02, Ajit Ghaisas <ajit.ghaisas at oracle.com <mailto:ajit.ghaisas at oracle.com>> wrote:
>>>>>>>>> 
>>>>>>>>> Hi Alexey,
>>>>>>>>> 
>>>>>>>>>    I tried building this patch with latest http://hg.openjdk.java.net/jdk/jdk/ <http://hg.openjdk.java.net/jdk/jdk/>
>>>>>>>>>  
>>>>>>>>>    1. Some basic copy paste errors are resulting in build failures
>>>>>>>>>    2. MTLTexturePool.m file is missing from the patch
>>>>>>>>> 
>>>>>>>>>    Can you please check & update?
>>>>>>>>> 
>>>>>>>>> Regards,
>>>>>>>>> Ajit
>>>>>>>>> 
>>>>>>>>>> On 30-Apr-2019, at 2:52 PM, Alexey Ushakov <alexey.ushakov at jetbrains.com <mailto:alexey.ushakov at jetbrains.com>> wrote:
>>>>>>>>>> 
>>>>>>>>>> Hello,
>>>>>>>>>> 
>>>>>>>>>> Here  is an update on our effort to use Metal framework for java2d rendering.
>>>>>>>>>> 
>>>>>>>>>> We’ve added image rendering support and some support for LinearGradient. Also, the code has been refactored.
>>>>>>>>>> 
>>>>>>>>>> Please have a look:
>>>>>>>>>> 
>>>>>>>>>> http://cr.openjdk.java.net/~avu/JDK-8220154/webrev.01 <http://cr.openjdk.java.net/~avu/JDK-8220154/webrev.01>
>>>>>>>>>> 
>>>>>>>>>> Best Regards,
>>>>>>>>>> Alexey
>>>>>>>>>> 
>>>>>>>>>> 
>>>>>>>>>>> Hello,
>>>>>>>>>>> 
>>>>>>>>>>> As far as we know Apple has deprecated OpenGL on MacOS platform (https://developer.apple.com/macos/whats-new/ <https://developer.apple.com/macos/whats-new/>).
>>>>>>>>>>> 
>>>>>>>>>>> Unfortunately, this decision greatly affects our products that based on Java Client technologies. So, we (here at JetBrains) decided to start a project to replace OpenGL rendering on MacOS platform with a new one based on Metal. This is a huge task, so we  decided to leverage current rendering architecture that is used in OpenGL rendering pipeline on Mac.  
>>>>>>>>>>> 
>>>>>>>>>>> That’s why we didn’t use MTKView for representing AWT windows (that probably would be a better approach in the long term). Currently we're using CAMetalLayer within AWTView. We’ve implemented flat color shape/curve rendering so far and there is a lot of work to do. So, we’re looking forward to any collaboration.
>>>>>>>>>>> 
>>>>>>>>>>> In the mean time I’d like to share our current work to discuss metal pipeline architecture at the early stage of work.
>>>>>>>>>>> 
>>>>>>>>>>> Here is the webrev with our on going development:
>>>>>>>>>>> 
>>>>>>>>>>> http://cr.openjdk.java.net/~avu/JDK-8220154/webrev.00 <http://cr.openjdk.java.net/~avu/JDK-8220154/webrev.00>
>>>>>>>>>>> 
>>>>>>>>>>> Best Regards,
>>>>>>>>>>> Alexey
>>>>>>>>> 
>>>>>>>> 
>>>>>>> 
>>>>>> 
>>>>> 
>>>> 
>>> 
>> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.java.net/pipermail/2d-dev/attachments/20190708/320c5931/attachment-0001.html>


More information about the 2d-dev mailing list