Maaajor performance problems with TSE

#21

Quote:this would obviously indicate that TSE might not yet be ready for prime time... or maybe i'm reading this wrong...

Well, that wasn't how I understood the original question, sorry! With an "average" estimate of a year from vision to release, starting a game now on TSE is absolutely viable. If you want to release a game to the public as a commercial project next month, then no, it's not "ready for prime time" on that schedule.

#22 Thx Stephen... that's what i was trying to find out... or at least get a feel for as far as the current status of the current TSE milestone for commercial use in the immediate timeframe (within a month or two)...

yes Brian, modding the prefs values to force ps/vs 1.1 do increase the framerate up to between 27 to 34... and movement is a lot smoother (as is the mouse)...

and, as expected, the shader effects are somewhat 'muted' at 1.1... bumpmapping in particular... but it still looks nice enough to run with...

thx again for the straight up info guys...

--Mike

#23 The Source Engine (HL2) also forces FX5200 and FX5600 cards to 1.1 in order to get any sort of decent performance out of them. nVidia really dropped the ball on the whole FX5XXX series (slow performance, flaky drivers, fans that sound like vacuum cleaners, etc). It really is best to avoid that line of cards altogether. If you are looking for something in that price range grab a Radeon 9600 Pro. Very, very good card with great performance and solid 2.0 shader support. It is what Ben is developing on so you know you won't go wrong there =)

#24 Thx for the follow up info Matthew...

yeah, the GeForce card was just to make sure that some other stuff i was working on ran on
both Radeon and GeForce cards... it will soon be relegated to the min requirement test machine :)

--Mike

#25 Hmmm, just got TSE today, thought it was running kinda slow so I checked the specs. I seem to be getting 10-20 fps in the new terrain water demo (800X600 windowed) on a P-M 1.6ghz with a 256mb geforce6800 go (pci-express). Just to test things out I dl'ed the car demo and got roughly 100 fps (800X600) which seems to be about right if a 6800GT got 150. I was wondering if the new terrain and water are supposed to be running this slow or if there's something wrong with the way I'm doing things.

Also, why is cpu skinning slowing the demo down to a couple frames per second? TGE does it fine with no slowdown it seems.

#26 Wow 10-20 fps ? I was seeing a friend run it at 30-50 fps on a Radeon 9000 mobile 64 mb (on a 2 year old laptop)... (with full terrain and water in view)

#27 Hmmm... the plot thickens...

#28 Are you guys compiling in release mode? Even then 10fps is really slow for a 6800.

#29 All my testing was on the downloadable demos - which I assume were release compiles.

#30 Nah, debug, I tried release, got it up to 30-40, 70ish when i wasn't looking at the water... still seems a bit slow though... anyone else with a similar system getting better speeds?

#31 @Eric, you might want to try removing the water from the scene to see if that gives you a boost. You can also try removing the terrain. If you remove both and still get a bunk framerate, something is very wrong. You can also try going fullscreen, this seems to help on nVIDIA cards.

#32 @Brian : I tried some metrics, Here are the results :
Atlas Terrain Only : 155 - 160 FPS
Water + Atlas Terrain : 102 FPS
Water + Atlas + 1 Shape (1000 faces) : 80 FPS
Water + Atlas + 1 Shape + fxEnvReplicator (1024 Grasses, 512 bushes, and ~30,000 foliages rendered) : 55 FPS
Water + Atlas + 1 Shape + fxEnvReplicator + Kork : 30-40 FPS

So if, I calculate :
Water takes : 53 FPS - 40.76 % of "free FPS" (considering free FPS are 25 to 155)
1 Shape : 22 FPS - 16.92 % "free FPS"
fxEnvReplicator : 25 FPS - 19.23% of "free FPS"
Kork : 25 FPS - 19.23% of "free FPS"

I thought FPS was directly impected by poly number. But it seems to be wrong. Rendeing 2 shapes of with total faces of 5,500 takes more time than 54,576 faces ! (1024*12 + 512*24 + 30000)

Is there something wrong with TSShapes ? or do I miss something?
I am realy surprised by kork model. It very slow down the whole things...

#33 Vincent, you're comparing very different things here.

1024 grasses means nothing. Are they animated? Do they receive lighting? Are they reflected somewhere? What resolution is the texture, and how many faces do each grass element have? Grass usually is transparent, a model is not.

You can't just say it's a grass and compare it with something that works fundamentally different. :p

Having said that, there's a line of improvements that will improve the rendering in TSE and from what I've heard they are coming soon.

#34 1024 grasses : they are animated, they recieve light (Sun), water reflect some of them, texture resolution : 512x512 bumped, with alpha , and 12 faces for each. And there is 8 different models of grass.

So if I understand, kork, does not use the same render function than TSStatics. I said I thought faces number was directly involved in FPS. But, with this example, I think there is some time issues elsewhere. But where ?

edit:

Quote:
Having said that, there's a line of improvements that will improve the rendering in TSE and from what I've heard they are coming soon.

Edit : Yep call of drawIndexed Primitive, grouped by materials as I understand last Brian's plan. But, he mentionned a 40 FPS for 10 kork models.

If there is some time issues in TSShapes, I think call to drawIndexed primitive won't cut this (these) time issue

#35 Poly count does directly affect frame rate, but so do a lot of other things. Changing shaders and shader constants can have a much larger affect on performance than sheer poly count. That's why the recent batching changes I've made improve the performance so much. There are many other factors though; the number of render state changes, shader size, number of texture units in use, size of textures, size of draw batches, etc.

#36 @Brian - I know its bad to ask when but... When are you posting the batching optimizations to CVS?

#37 I think he said in his last plan after Ben is back from vacation so he can add his Atlas changes in as well as any last minute stuff.

From Brian's last plan dated June 6 2006

Quote:This update should go live pretty soon, I'm just cleaning up some minor remaining issues and waiting for some Atlas changes to be complete before it's released.

#38 Found NIVIDIA 7500 LE to be slower on OpenGL than D3D on Windows vista. May be worth switching to D3D and see if there is an increase in speed on TGE.

Edit: Changed default MS driver to manufacturers - resolved my problems.