www.jpct.net
jPCT-AE - a 3d engine for Android => Support => Topic started by: EgonOlsen on December 17, 2010, 11:20:01 pm
-
I'm thinking about doing a version of An3DBench for faster devices, but i don't have access to one. I need to know how fast a current device actually is under load to tailor the tests accordingly. I've created a little terrain hack and would be interested in the frame rate. So if anybody with a fast device is willing to test it, i would be very grateful. Here it is: http://www.jpct.net/jpct-ae/download/tmp/An3DBenchXL.apk (http://www.jpct.net/jpct-ae/download/tmp/An3DBenchXL.apk)
It loads a terrain and runs forward in a straight line in fixed intervals, i.e. no timing or even an end (just exit the test with the home button). It logs fps to the console and i would be very interested in three things:
- the hardware/Android version being used
- does it look ok?
- and finally: the fps!
Here's how it is supposed to look (including z-accuracy problems that occur on my phone):
(http://www.jpct.net/jpct-ae/download/tmp/terrain.png)
On my Samsung Galaxy with 1.5, it runs @ 2fps...i expected less, so it's actually not that bad... ;D
-
I tested it, but I dont know how can I gen some results... console? what is it? :D ...
-
The DDMS logcat tab in Eclipse shows it for example. Or, if you are on the phone only, Android System Information (available in the market) can display it.
-
* stock N1 / Android 2.1
* it looks mostly ok. there are some tiny vibrations on some parts, possibly related with z-accuracy thingy. but mountains look gray as opposed to green ones in your screenshot
* starts with 16fps and increases up to 32 fps (when nothing seen)
here is a shot. dony worry about vertical lines. it happens since eclipse cant capture screen atomically
(http://img571.imageshack.us/img571/7231/jpctbenchmark.png)
-
Performance is ok, but it looks like as if your phone/driver doesn't care about vertex alpha...too bad...
-
On Galaxy S with custom tweaked 2.2.1 starts with 38fps and 43fps when is nothing to see... looks be fine :)
-
Fine means that there is some green visible? On which kind of device?
Edit: Sorry, i somehow missed that Galaxy S part...
-
Performance is ok, but it looks like as if your phone/driver doesn't care about vertex alpha...too bad...
maybe updating to 2.2 fixes it. i'll retry it when i upgrade (dont now when)
-
You are very fast, I edited my post... It looks same as your screenshot :)
-
On my Droid X, Android 2.2, it runs at a steady 21-22 fps, and looks gorgeous (edit: that is when the mountains are visible, goes up a couple fps once nothing is visible, of course)
-
Looks like as if the Nexus One has a problem with vertex alpha while other phones (including my sucky Samsung) work fine. That's pretty sad, because i can't use this feature in this test then, which will make the terrain look much more dull. I will increase the polygon count instead and add some trees, i think...
-
Updated the test with a version that doesn't use vertex alpha but some lens flares and lighting effects instead. Does that work on Nexus One? It works fine on my phone except for some z-fighting in the water which creates a kind of saw tooths at the coasts...but i can live with that.
-
On galaxy S with 2.2.1 it is running about 42fps, but I think that textures of lenses are bad... scene is nice :)
(http://www.freeimagehosting.net/uploads/6ac7e63d9f.jpg)
-
The textures are actually fine, but the rendering isn't that good. The gpu somehow thinks that dithering the black parts is needed...i've no idea why, it's the same on my phone. I'll see if i can fix this...
-
seemed ok to me. but not sure?
(http://img507.imageshack.us/img507/5807/jpctbenchmark2.png)
-
Looks fine. Your water has the same accuracy issues as mine, the Samsung Galaxy S obviously doesn't suffer from those. But that's no problem IMHO. I'll upload a version with a real benchmark path and trees soon.
-
Updated the APK again with the mentioned changes. It's a "real" benchmark run now with a result at the end. On my phone, performance is down to 1fps (giving me 1000 points... ;) !!
-
fps 13.77
score:13770
-
Many thanks. So the Nexus One is 13 times faster in this test than my phone...can't wait to see some Galaxy S results.
BTW: Today i saw some An3DBench results from the Nexus One under 2.3. It's quite an improvement when compared to 2.2 (around 20%)...nice...
-
Galaxy S : 26,07 / 26070
-
Wow! 26 times faster than my phone...i'm about to cry... ;)
Edit: For the record:
Samsung Galaxy: 1.01 fps / 1010 points.
-
How many polygons are rendering? Only 26 fps on galaxy S, it is not good ;D
-
Around 140.000 triangles/frame (not all are visible of course).
Edit: triangle count is only one part of the equation, the other one is object count. The test renders the water and the terrain in large chunks, but each tree is an individual object. The test has 500 trees visible (out of 3000 for the whole terrain) at a time at max. Has anybody noticed some tree popping, i.e. trees that are there in one frame but disappear in the other? That might happen if the number of visible trees exceeds the tree pool size of 500...but i can't try this for myself as my frame rate is so low.
-
Droid X, Android 2.2 results:
Terrain rendering: 13.82 fps
TOTAL SCORE: 13820
-
i havent noticed any flopping
-
Me neither...i had the possibility to run it on a Galaxy S. Lens flare textures still look bad though...i don't really know what to do here. There's no point in dithering black areas. I've no idea why it's doing this...
-
HTC Desire 2.2
32bit, VBO on:
14.47fps
score 1440
Absolutely the same results w/ 16bit, VBO off o_O
Water looks ugly in parts where it intersects w/ terrain. I noticed this hardware issue in my apps as well.
-
Yes, it's an accuracy issue with the zbuffer on Qualcomm chips. There's nothing i can do about this, so i'll simply live with it.
-
Sorry for bothering you all again, but i need more feedback... ;D I've replaced the terrain test with another one on which i need some performance feedback. This time it's (unlike the terrain) not very gpu intense, but very cpu/vm intense. On my phone, it runs @2+fps...again.
Screen shot of how it's supposed to look like:
(http://www.jpct.net/jpct-ae/download/tmp/cloth.png)
Download: http://www.jpct.net/jpct-ae/download/tmp/An3DBenchXL.apk (http://www.jpct.net/jpct-ae/download/tmp/An3DBenchXL.apk)
-
39,58fps on galaxy S with 2.2.1... very nice test ;)
-
Holy shit...almost 40fps. I wasn't expecting a frame rate that high. Might be the JIT and the fpu that really help here. Thanks for testing.
-
~13-14 fps on N1
-
I hope to see a test of new generation Qualcomm's processors that have Desire HD or G2 :)
-
~13-14 fps on N1
With 2.1...with this test, a run with 2.2 would be really interesting to see if the JIT can do any good.
-
Droid X, Android 2.2 results:
Terrain rendering: 42.64 fps
TOTAL SCORE: 42640
Edit: I did several runs to get a range of scores. It seems to range between 40 - 42 fps (fastest score posted above). Very cool animation, BTW!
-
Cloth animation benchmark 54,79 FPS ;D
Phone: Desire HD
Stock HTC Rom, Android 2,2
-
Cloth animation benchmark 54,79 FPS ;D
Phone: Desire HD
Stock HTC Rom, Android 2,2
Just *WOW*.
-
Cloth simulation
HTC Desire stock 2.2
33.19 fps
Total score: 33190
-
I've updated the benchmark with a version that includes all three tests (added one that uses raft's bones and some vertex controller stuff) and gives a "proper" result at the end, i.e. the score is actually calculated correctly. To make you envious of me and my phone (Samsung Galaxy, Android 1.5):
Skeletal animation: 2.01 fps
Cloth simulation: 2.15 fps
Terrain rendering: 1.00 fps
TOTAL SCORE: 2051
BTW: This version finally prints out the results to the console, so that you can simply copy them to post.
I'm still looking for one or two additional tests. Any ideas?
Edit: Screen shot of the new test:
(http://www.jpct.net/jpct-ae/download/tmp/ninja.png)
-
result of Galaxy S with 2.2.1 ...
Skeletal animation: 15.30 fps
Cloth simulation: 40.39 fps
Terrain rendering: 31.67 fps
TOTAL SCORE: 34738
-
Stock N1, 2.1
Skeletal animation: 9.19 fps
Cloth simulation: 12.68 fps
Terrain rendering: 13.92 fps
TOTAL SCORE: 14231
why is skeletal animation so slow? Bones demo app runs @ 50+ fps with one ninja. is the scene a high poly one or ?
edit: Bones demp runs 40+ fps with animation, 50+ is without animation, stupid me ::) but still too slow compared to benchmark
-
I tested bones every with 8 ninja and I have 55fps without animation, 26 with skin animation and 54 with mesh :)
-
I tested bones every with 8 ninja and I have 55fps without animation, 26 with skin animation and 54 with mesh :)
26 fps with 8 ninjas, impressive ;D your demo app might be slightly outdated, i remember as i removed mesh animation from demo
-
The animation test is more like an alpha blending, high polygon, high object count and animation test. The animation itself runs pretty fast, the test suffers most from the high polygon vount of the structure (22,000 iirc) and lots of individual, alpha blended grass objects. I think i'm going to rename it and give it an abstract name instead.
-
I think i'm going to rename it and give it an abstract name instead.
i will be glad. now it gives the impression that skeletal animation is too slow
-
I think that i'm going to remove that structure. It doesn't look as demanding as it is, which makes people wonder where the performance goes...as we can see.. :)
My initial idea waa to add something that looks a little bit asian, but free resources were pretty limited, so i took this one. But i guess i'll just go with some trees and maybe extend the grass' animation instead.
-
if you send me some picture, what you want modeling, I can do it :) ... but I am not big friend with texturing part
-
HTC Desire HD:
Skeletal animation: 16.31 fps
Cloth simulation: 52.75 fps
Terrain rendering: 32.06 fps
TOTAL SCORE: 40210
HTC Hero:
Skeletal animation: 2.30 fps
Cloth simulation: 3.07 fps
Terrain rendering: 1.18 fps
TOTAL SCORE: 2604
my netbook running froyo x86 :P
(http://i52.tinypic.com/vhxdw.jpg)
Suggestions:
- a scene with more animated objects
- a scene with some particle effects
-
I've updated the test with a version that doesn't use the structure in the animation test anymore. No idea if that's much faster, performance may still suffer from lots of animated grass. In addition, i changed the tests' names to something more...prosaic. The score calculation has been changed to compensate for the missing structure. Base line is to have 2000 points on my crappy phone.
I'll add a particle test, but no test with more animations because of two reasons: a) Missing art work and b) the Ninja test already features dozens of animated objects, even if they are pretty small (the grass!).
-
Stock N1, 2.1
Ninjas' garden: 10.57 fps (slight increase, last 9.19 fps)
Emperor's new clothes: 12.38 fps
Magic island: 13.75 fps
TOTAL SCORE: 12112
is grass animated via Bones?
btw, IMHO skeletal animation test make more sense and look more interesting if it consists of several ninjas performing different animations.
-
Galaxy S with 2.2.1 ... look at skelet and ninjas :D ... and what add test same as cloth but computing by GPU?
new:
Ninjas' garden: 15.59 fps
Emperor's new clothes: 40.79 fps
Magic island: 32.07 fps
TOTAL SCORE: 29191
old:
Skeletal animation: 15.30 fps
Cloth simulation: 40.39 fps
Terrain rendering: 31.67 fps
TOTAL SCORE: 34738
-
No, grass' animation is done by using is a simple vertex controller. However, each patch has 88 vertices and there are 72 patches in this scene, so there's quite a lot stuff to handle. I thought about playing different animations, but i would end up showing multiple ninja's making the same ridiculous moves over and over again. While this might be a good example of skeletal animation from a techy stand point (and it's basically what i did in the former version of An3DBench with the alien), it doesn't look too great IMHO. A kind of fighting scene would be much cooler, but i neither have to the time nor the talent to create it. That's why there's a ninja walking around scrunching grass now.
-
... and what add test same as cloth but computing by GPU?
Well...yes...No! That would require OpenGL ES 2.0 support, which means that a) i can't test it due to lack of hardware and b) a lot of devices won't be able to run it. In addition, the cloth simulation is pretty complex especially as it can interact with any geometry in the scene, which means that it requires collision detection with client side geometry...something that's always a problem when doing physics on the gpu (which is why most games using PhysX on the gpu are limiting it to eye candy that doesn't really interact with the scene like debris and similar). I'm also not sure if ES 2.0's shader support would be sufficient to do it. Apart from that, it would be a huge amount of work, even if all preconditions were met. So, as said: No, that won't happen.
-
Uploaded a version with a basic particle test. It's based on the particle "system" that i wrote for Alien Runner, so it's pretty optimized. It renders up to 500 particles...and i have no idea if it looks good in motion, because i have no real motion on my phone (around 3fps). Here's a screeny:
(http://www.jpct.net/jpct-ae/download/tmp/flowers.png)
-
Opps, forgot to upload the actual apk...stand by...
-
Upload is fine now....stupid me..
-
Flowers? ;D
Galaxy S 2.2.1 ... 22.31fps
-
Nexus One, 2.1
Flower power: 11.87 fps
-
Flower power: 16.38
Ninjas garden: 15.48
clothes: 43.51
Magic island: 29.66
Score: 22733
Egon, have you already thought about installing a custom rom on your galaxy? With froyo you will have more fps.
http://forum.xda-developers.com/showthread.php?t=865342 (http://forum.xda-developers.com/showthread.php?t=865342)
-
Egon, have you already thought about installing a custom rom on your galaxy? With froyo you will have more fps.
http://forum.xda-developers.com/showthread.php?t=865342 (http://forum.xda-developers.com/showthread.php?t=865342)
I did...and decided against it for the time being. Judging from the results of other, similar phones running Froyo, i don't think that it'll be that much better. 2fps compared to 1.5fps is still slow. These phones suffer more from the absence of an fpu than from Dalvik's poor performance.
-
Uploaded a new version with a new test. With that one, i've five tests in place now and i'll leave it at that for now. The next step is to think about the option to upload and compare results...
The new test isn't much more than an extended version of the rotating beethoven in "normal" An3DBench. This time it's a dragon with a higher polygon count and some nice (fake, with incorrect lighting...) reflections. Note that the gaps don't reflect the dragon but the stones do (it's actually a cheap trick, but i liked it... ;)).
(http://www.jpct.net/jpct-ae/download/tmp/dragon.png)
-
Hey, you said to me, that reflections are almost impossible in OGL1, and now I see very nice reflection in your benchmark :) how can you do it? These are two dragons? ... so and result...
Double dragon: 31.89 fps
Flower power: 23.03 fps
Ninjas garden: 14.87 fps
Emperors new clothes: 40.95 fps
Magic island: 31.34 fps
TOTAL SCORE: 24602
-
Desire HD:
Double dragon: 42.16 fps
Flower power: 15.12 fps
Ninjas' garden: 14.58 fps
Emperor's new clothes: 44.51 fps
Magic island: 28.70 fps
TOTAL SCORE: 25120
Nice. Maybe you could write a small example of this trick :D
-
As said, it's pretty much fake. It's the same mesh that's used two times in the scene. The second time with inversed culling and a tweaked rotation matrix (set to -1 at 1,1). There's a transparent plane between them to fake the floor. As said, lighting isn't correct with this approach. The mirrored dragon should actually be lit like the original one, but instead it's lit on its own. It should be possible to get this correctly by really seperating the two into two different worlds and rendering them both into one buffer, but for this benchmark, i haven't done it. The human eye is very forgiving concerning wrong lighting, so i think i can get away with this.
-
nice reflections and nice trick ;D
Nexus One 2.1
Double dragon: 11.10 fps
Flower power: 11.72 fps
Ninjas' garden: 9.46 fps
Emperor's new clothes: 12.18 fps
Magic island: 13.83 fps
TOTAL SCORE: 10093
-
In all versions of this an3Dbench XL I must double tap on VBO or textures to turn on...
-
32bit and vbos are both enabled by default. The gui just doesn't reflect this. I'll fix this.
-
maybe you can disable possibility to change these items, for better comparison
-
Droid X, Android 2.2 results:
Double dragon: 15.97 fps
Flower power: 20.42 fps
Ninjas' garden: 11.39 fps
Emperor's new clothes: 43.72 fps
Magic island: 17.38 fps
TOTAL SCORE: 18853
-
maybe you can disable possibility to change this items, for better comparison
I agree. I'll remove them.
-
HTC Desire 2.2
Double dragon: 11.49 fps
Flower power: 13.76 fps
Ninjas' garden: 9.30 fps
Emperor's new clothes: 35.25 fps
Magic island: 15.14 fps
TOTAL SCORE: 14708
-
Egon, have you already thought about installing a custom rom on your galaxy? With froyo you will have more fps.
http://forum.xda-developers.com/showthread.php?t=865342 (http://forum.xda-developers.com/showthread.php?t=865342)
I've finally updated the phone (It's an original Galaxy, not a Spica...so the mod in the link doesn't apply btw.) to Galaxo 1.6.3.4 with a backported Froyo VM (i.e. with JIT). It's noticable faster in everyday usage, but not in my 3D benchmarks. They are either limited by the GPU or the absence of a fpu, the jit does no good in either of these situations.
-
It's too bad. I thought it would help.
But if the normal apps are running faster it wasnt waste of time.
Btw can you please look into my last topic in Support please. ::)
-
Btw can you please look into my last topic in Support please. ::)
That question with centering the text? if so, i've no idea. I've never used this stuff myself.
-
Viewsonic 10'' Gtablet Android 2.2
Results(1024*600)
Double dragon: 21.99 fps
Flow power: 29.47 fps
Ninjas' garden: 18.98 fps
Emperor's new cloths: 51.54 fps /* cloth did not seem to show texture, was all black */
Magic island: 25.50 fps
TOTAL SCORE: 25537
-
Emperor's new cloths: 51.54 fps /* cloth did not seem to show texture, was all black */
That's a bit strange...maybe a problem with the normals....i'll have a look.
-
I've updated the version in the market. I wasn't calculating normals for the cloth..i've changed this...i hope this helps.
-
"Package file was not signed correctly."
-
Uninstall any version downloaded from this site before installing the market version. Android has problems mixing them somehow.
-
nice result (http://www.jpct.de/an3dbenchxl/details.php?code=ODIyMwxx157410) of tegra 2
-
CPU performance is really high, but graphics performance itself isn't that great. It only wins because it totally dominates in the clothing test.
-
Benchmark ran smoothly on HTC Desire running the latest test build of Oxygen ROM.
CPU usage was around 64% - 67%
http://www.jpct.de/an3dbenchxl/details.php?code=MTE3OTMx66779 (http://www.jpct.de/an3dbenchxl/details.php?code=MTE3OTMx66779)
Any chance of release the source for the benchmark? :P