Flash 10 speed tests
May 28th, 2009
Sometimes coming from only one side of a problem restricts your view on the best way to evolve. In recent experiments with the latest Flash10 player, we have been testing how starting from a native cs4 approach fares when constructing a 3d framework, and have discovered a few interesting facts about Flash 10’s renderer. Some of these nuggets include the following:
- DisplayObject sorting is a much faster option than triangle sorting.
- Fixed length Vector data stores are faster than linked lists when it comes to manipulating large data sets.
- Sorting using index references is much faster (and easier) than ordering screenvertex vectors
- …
The above demo is one such test, and nicely illustrates where we feel Away3d is going after the textfields/shapes merge that will be version 2.4. Click to add more globes – speeds vary between machines but it seems most that have tried the test can manage up to 20,000 unsorted triangles, still running at 30fps.
Moving you mouse to the right or left side of the screen will switch the current z-sort algorithm on or off, and you can see with higher numbers of triangles there is a significant performance drop when sorting is activated. This uses a brute-force method – in more recent tests it seems a better approach is to sort in convex groups, and we will hopefully be able to post more on this soon.
The Away3d Team
Popularity: 77% [?]
Hi,
Look promising :)
Colud you explain though why Away 3D FP10 branch is so much slower
compared to PV3D which make no use of new drawing API ?
I experienced big drop in performance im my own project when using Away3D ( no offence guys, just facts ) and it’s also well tested here – http://clockmaker.jp/blog-en/2009/05/away3d_vs_pv3d/.
I suspect that it’s connetced with perspecitve correct bitmap distortion
which is actualy slower than affine transformations even if it looks pretier and helps to lower
number of polygons on flat surfaces.
Am I right ? Don’t you think that it woluld be desirable feature to have an opt-out from perspetive corrected drawing
to pre FP10 way ? Then one could optimize the scene better – apply perspective correction to floor and walls but
use faster afine transforms for objets/surfaces that don’t really need perfectly accurate texture rendering ?
Thumbs up for the great work !
Comment by jkozniewski
— May 29, 2009 @ 2:06 am
18 FPS on 50160 Faces with ZSort = false
AMD Athlon 64 X2 Dual 2.71Ghz, 2GB RAM
Comment by Exey
— May 29, 2009 @ 5:40 am
@jkozniewski
the speed tests on the clockmaker blog are a little biased, because the flash10 drawtriangles in it’s basic form actually slows down the rendering process. If you were to compare the flash10 branch of papervision with the flash10 branch of Away3d, you would see a much closer speed comparison.
The above demo illustrates where we are coming from with the next major flash10 update. Keeping current features in Away3d means that ultimately we can’t hit 20,000 tris at 30fps, but we are testing (with some success!) the possiblity of turning off features like interactivity and sorting in order to allow maximum speed for those that require nothing else.
Comment by Rob Bateman — May 29, 2009 @ 9:29 am
@Rob Bateman
Ok I understand ;) I was just a bit ( unplesantly ) supprised that FP10 branch is actually slower than
older ( FP9 ) engines and your explanation seems to confirm my suspicion that new features
like drawTriangles with UVT are not always the fastest way to render vast number of polygons.
I just wonder what’s “not basic” form of drawTriangles that could outperform old ways of drawing ?
Whats more… colud please comment on idea of flexible renderer that could mix new and old
way of drawing ( since it seems that in some cases the old, I mean affine transforms here, way is faster and sufficient in terms of apperance ).
Comment by jkozniewski
— May 29, 2009 @ 11:20 am
@jkozniewski
Well, drawTriangles can be pretty fast as you can see from the above demo, if used correctly. The reason why this is not currently the case in the flash10 branch of both Away3d & Papervision, is that the old render techniques separated each triangle draw routine into it’s own method call. drawTriangles works best when a whole bunch of triangle data is rendered at once, and this requires some fairly hefty re-writing of internal data structures as you can imagine. However, we expect to be looking at the options very soon, and will of course update everyone with progress when we have more.
Comment by Rob Bateman — May 29, 2009 @ 2:16 pm
20 FPS on 50160 Faces | ZSort = false
AMD Phenom II X4 940 | 3 GB RAM
Comment by DonPaolo — May 30, 2009 @ 7:46 am
@Rob, in other words, you are not going to keep affine rendering in any way, right? I heared older tests found no significant difference between series of drawTriangles calls and single call performance, has this changed now?
Comment by makc — June 22, 2009 @ 4:33 pm
oh yeah, the results,
in high quality, 6840 faces run at 23(31) fps with zsort on(off)
in low quality, 53200 faces run at 3(4) fps with zsort on(off)
(feel compelled to point out that sandy was sorting in groups since forever, whoever came up with that idea – Thomas? – is a smart guy)
Comment by makc — June 22, 2009 @ 4:42 pm
mesh:100
face:76000
z-sort:false
fps:12-18
mem:56mb
z-sort:true
fps:5-18
mem:61mb
Great!!
Comment by zszen
— June 29, 2009 @ 4:56 am
Honestly I’ve been digging around looking into all of the flash 3D engines and to me away 3D looks the best, might arguably not perform the best RIGHT now but I think is the most promising. I’ve messed around with papervision and at that time didn’t like my scenes freezing every 5-10 seconds getting choked up refreshing…. don’t know how much better it is now but..
Anyway with 28880 faces amd X2 64 (on 32 bit plat) 3 gigs ddr2-800 getting about 20 fps avg.
Comment by Jesse — July 21, 2009 @ 1:09 pm