[pending]irr transplant to android with float optimization

[混江龙wade]

transplant the engine to android and float point optimization (using SIMD),but only found no performance improvement,i searched the Internet and found a technical post(http://bbs.meizu.com/viewthread.php?tid=1903677&extra=&page=1) which says they optimized the float-point performance and the performance increase by 3 times ,tried to contact the owner of the post but without success,If it is 3 times performance improvement on the pc float technology,I think it is very easy to think of,Therefore,we feel that perhaps this floating-point improvement has nothing to do with irrlicht itself ,but with android,i really want you answer ,sorry for my poor english.

[CuteAlien]

Your link does not work.

[aaammmsterdddam]

err, yeah it does. check it again

*here it is in case you selected wrong*

http://bbs.meizu.com/viewthread.php?tid ... ra=&page=1

[hybrid]

Interesting, now it does work. Anyway, it would be interesting for which system you develop. Is it the Meizu as mentioned in the forum? Anyway, I doubt that you get any major performance increase for a generic app. At least not if you have a proper 3d acceleration and don't do any software rendering which you want to improve. What might be optimized is the animation system. But I'd suggest to do a profiling of a typical scene first, and check if you don't have much more potential on the render side. Make sure that the node sorting is adopted to the GPU and that the necessary image sizes are used.

[hendu]

@hybrid if you're interested, the last I profiled irr the top cpu user was the matrix math (makeInverse and multiplying two matrices). Neither was high enough to matter though.

[混江龙wade]

@CuteAlien
thanks,it seems work now ,please try it again
http://bbs.meizu.com/viewthread.php?tid ... ra=&page=1

[混江龙wade]

@hybrid Meizu is a mobile phone brand which used android,I am a student in East China university of science and technoloy,never did I perform an optimization for any app.I have profiled some scene with Intel VTune Amplifier and found it's hard to get any improvement on matrix multiplication and vector product. The LBE Team work for Meizu,They claimed that the floating-point performance increase three times and illustrated with the example helloworld ,I know helloworld is not a complicated example.yesterday I found another post ,does it help to my optimization?
http://hilbert-space.de/?p=22
http://blogs.arm.com/software-enablemen ... plication/

[Virion]

by the way saw this on the meizu forum:

Gameloft正是用irrlicht来开发iPhone游戏的

Translate: Gameloft uses Irrlicht for developing iPhone games.

o__O WHATTTT never knew that

[混江龙wade]

@Virion This is not very important,what they want to say is that Gameloft is trying to develop games using Irrlicht.

[fmx]

by the way saw this on the meizu forum:

Gameloft正是用irrlicht来开发iPhone游戏的

Translate: Gameloft uses Irrlicht for developing iPhone games.

o__O WHATTTT never knew that

Yeh I noticed this when Sky-Fox uploaded his videos here viewtopic.php?p=266255#p266255
see what banner loads up at 0:40 in the first video?

personally I dont know why GameLoft want to make the switch, but I guess this is only a good thing for the future of irrlicht, provided they give credit or mention it


On topic:
The results will vary from application to application, depends entirely on how the math functions get used:
the more frequently an app uses heavy maths, the bigger the performance gains

Meizu must have done other optimizations on top of floating point math, which they either dont want to talk about or forgot to mention.

I dont believe using float-optimized maths with standard irrlicht hello-world demo would be enough to result in such a big performance gain, they definitely made other changes to irrlicht source

[qidaozhilong2]

You can try to add "-ffast-math" to your makefile
It's gcc float compiler command.
Anyway, i think this will increase your FPS performance, but actually it's not safe.

[REDDemon]

it's not IEEE compliant. and maybe has a bit less precision. if it is not compliant probably you can't send those float to the GPU .

http://gcc.gnu.org/ml/gcc/2001-05/msg01566.html