[OpenJDK 2D-Dev] [9] Review Request: 8029253 [macosx] Performance problems with Retina display on Mac OS X

[Sergey Bylokhov]

Hi, Phil.
New CR were filed:

JDK-8059941 [D3D] The fix for JDK-8029253 should be ported to d3d pipeline
JDK-8059943 [macosx] Aqua LaF should use BI.TYPE_INT_ARGB_PRE for a 
better performance
JDK-8059944 [OGL] Metrics for a method choice copying of texture should 
be improved
JDK-8059942 Default implementation of DrawImage.renderImageXform() 
should be improved for d3d/ogl

On 08.10.2014 3:34, Phil Race wrote:
> Just got a chance to run the performance test I'd say we've got 
> exactly the same penalty
> in D3D .. and hopefully a similar fix will help.
>
> -phil.
>
> On 10/7/2014 12:57 PM, Phil Race wrote:
>> Looks good. I don't have a PC handy right now to test myself but I'm 
>> curious if the
>> tests has similar poor performance issues with D3D and whether adding 
>> the
>> same would help there too ..
>>
>> -phil.
>>
>> On 10/6/14 10:29 AM, Sergey Bylokhov wrote:
>>> Hello.
>>> Please review the fix for jdk 9.
>>>
>>> The new blit was implemented: OGLGeneralTransformedBlit, it adds a 
>>> huge boost in case of uncached/unmanaged images:
>>>
>>> Bug: https://bugs.openjdk.java.net/browse/JDK-8029253
>>> Webrev can be found at: 
>>> http://cr.openjdk.java.net/~serb/8029253/webrev.04
>>>
>>> Notes:
>>> - The fix will help only if a non-ARGB_PRE images with alpha or 
>>> getdata() were used. I'll fix such usage in swing in separate CR.
>>> - i have discarded the change from the previous version of the fix: 
>>> http://cr.openjdk.java.net/~serb/8029253/webrev.00. It adds 30% 
>>> improvements in swingmark-native, but decrease performance on 
>>> win/lin. So I have decided to fix it separately.
>>> - Initially I used incorrect source offsets in the fix, which were 
>>> not found by the tests, so I add a new one. Commented types in the 
>>> test still fails as before the fix, I will check them 
>>> later.(probably related to JDK-6593406)
>>> - OGLAnyCompositeBlit for OpenGLTexture was removed, because it 
>>> tries to copy destination(texture), and such blit(texture -> 
>>> IntArgbPre) is absent.
>>>
>>>
>>> Summary osx-retina Intel HD Graphics 4000 :
>>> http://cr.openjdk.java.net/~serb/8029253/perf.04/osx-retina-intel/results.txt 
>>>
>>>   base:
>>>     Number of tests:  24
>>>     Overall average:  92015.36146212014
>>>     Best spread:      0.0% variance
>>>     Worst spread:     0.0% variance
>>>     (Basis for results comparison)
>>>
>>>   fix:
>>>     Number of tests:  24
>>>     Overall average:  616682.7498394073
>>>     Best spread:      0.0% variance
>>>     Worst spread:     0.0% variance
>>>     Comparison to basis:
>>> *      Best result:      4899860.24% of basis*
>>>       Worst result:     100.48% of basis
>>>       Number of wins:   22
>>>       Number of ties:   2
>>>       Number of losses: 0
>>>
>>> Summary osx-retina Nvidia GeForce GT 650M 1024 MB :
>>> http://cr.openjdk.java.net/~serb/8029253/perf.04/osx-retina-nvidia/results.txt 
>>>
>>>   base:
>>>     Number of tests:  24
>>>     Overall average:  147138.4952051871
>>>     Best spread:      0.0% variance
>>>     Worst spread:     0.0% variance
>>>     (Basis for results comparison)
>>>
>>>   fix:
>>>     Number of tests:  24
>>>     Overall average:  634465.9854992364
>>>     Best spread:      0.0% variance
>>>     Worst spread:     0.0% variance
>>>     Comparison to basis:
>>>       Best result:      42311.89% of basis
>>>       Worst result:     77.93% of basis
>>>       Number of wins:   16
>>>       Number of ties:   7
>>>       Number of losses: 1
>>> -- 
>>> Best regards, Sergey.
>>
>


-- 
Best regards, Sergey.