[OpenJDK 2D-Dev] RFR: 8251854 [macosx] Java forces the use of discrete GPU

[Philip Race]

On 8/25/20, 12:27 PM, Sergey Bylokhov wrote:
> On 25.08.2020 05:43, Kevin Rushforth wrote:
>> Does this only apply when the MacBook is running on battery, or will 
>> this affect performance even when the laptop is plugged in? If the 
>> latter, I wonder what Apple's rationale is for including a discrete 
>> graphics card that isn't used most of the time.

Based on the numbers, I wonder if we should make this change ?
>
> It is applied if the "automatic graphics switching" is enabled, if the 
> user disables
> this feature for the "power adapter" mode, then the discrete graphics 
> will be always used.

That's a bit misleading
If I disable automatic graphics switching it is disabled for BOTH batter 
and power
and vice versa. In other words there is no way to express that battery 
power should fall back
to integrated and that you only want discrete when running on the adapter.

-phil



>
> I guess by default they try to "maximize battery life":
> https://support.apple.com/en-us/HT202043
>
>>
>> -- Kevin
>>
>>
>> On 8/24/2020 11:27 PM, Sergey Bylokhov wrote:
>>> On 24.08.2020 13:35, Philip Race wrote:
>>>>
>>>> Is there any performance cost to doing this ? I'd expect so. Any 
>>>> estimate ?
>>>
>>> Yes, performance is affected for sure:
>>>
>>>  - SwingMark:
>>>    OGL_Base: 14000
>>>    OGL_Fix: 24000
>>>    Metal: 18000
>>>
>>>  - Here is a j2dbench for the common draw operations(new/old/metal):
>>>    http://cr.openjdk.java.net/~serb/8251854/perf/results.txt
>>>
>>>   Summary:
>>>   OGL_base:
>>>     Number of tests:  24
>>>     Overall average:  4.556306150323041E8
>>>     Best spread:      0.16% variance
>>>     Worst spread:     4.68% variance
>>>     (Basis for results comparison)
>>>
>>>   OGL_fix:
>>>     Number of tests:  24
>>>     Overall average:  1.0086929824044746E8
>>>     Best spread:      0.04% variance
>>>     Worst spread:     7.89% variance
>>>     Comparison to basis:
>>>       Best result:      83.41% of basis
>>>       Worst result:     15.73% of basis
>>>       Number of wins:   0
>>>       Number of ties:   0
>>>       Number of losses: 24
>>>
>>>   metal:
>>>     Number of tests:  24
>>>     Overall average:  8.841681616575797E7
>>>     Best spread:      0.08% variance
>>>     Worst spread:     5.64% variance
>>>     Comparison to basis:
>>>       Best result:      248.11% of basis
>>>       Worst result:     19.1% of basis
>>>       Number of wins:   8
>>>       Number of ties:   2
>>>       Number of losses: 14
>>> ==========================================
>>>
>>>  - Here is a j2dbench for the common draw operations(newOGL vs metal 
>>> only):
>>> http://cr.openjdk.java.net/~serb/8251854/perf/newOGL_vs_Metal.txt
>>>     Summary:
>>>   OGL_fix:
>>>     Number of tests:  24
>>>     Overall average:  2.5871177969681844E7
>>>     Best spread:      0.04% variance
>>>     Worst spread:     7.01% variance
>>>     (Basis for results comparison)
>>>
>>>   metal:
>>>     Number of tests:  24
>>>     Overall average:  2.1896134898150157E7
>>>     Best spread:      0.04% variance
>>>     Worst spread:     1.98% variance
>>>     Comparison to basis:
>>>       Best result:      488.31% of basis
>>>       Worst result:     30.77% of basis
>>>       Number of wins:   14
>>>       Number of ties:   0
>>>       Number of losses: 10
>>>
>>>> And there's then no way to explicitly request the discrete card on 
>>>> a 15/16" MBP.
>>>
>>>   I have checked that the discrete card is enabled by the macOS:
>>>   - if the full screen window is set
>>>   - if the second monitor is connected
>>>   Do not know any other ways to enable it.
>>>
>>>
>>>> Should we release note this ?
>>>
>>> Yes, I think so.
>>> Note that it does not affect the bundled applications only apps 
>>> running via java launcher.
>>> But some(most?) bundled java applications use this flag already.
>>>
>>>
>>
>
>