PDA

View Full Version : Is it me or ...



Musart
07-31-2006, 06:54 AM
I'm wondering why 9.0 Volumetric Light is much slower than 8.5 (I mean 5 to 10 times slower !!!)

I'm I missing something ?

P.S : Same scene and same settings used !!!

Mus

jameswillmott
07-31-2006, 07:23 AM
You should post the scene for others to confirm....

Musart
07-31-2006, 08:27 AM
You should post the scene for others to confirm....

You're right :agree:

Musart
07-31-2006, 04:20 PM
As suggested:

Here are the files (Scene and object)

Just load the scene, choose object and press F9

Compare 8.5 to 9.0


I'll post my results further tonight.

Mus

Musart
07-31-2006, 05:51 PM
Ok ...

Here are my results :

LW 8.5 : 56 sec

and mhmhm ...

LW 9.0 : 3330 sec (55m 30s)

So what's wrong ???

If someone can help me that would be Grrrrreat :-)

MUS

Lewis
07-31-2006, 06:13 PM
Confirmed, LW 8.5 renders this scene in 57 secs (on my machine while surfing and doing some work what uses only 5-6% of CPU time - 3GHz P4) while LW9 (on both cameras - classic/perspective) is much, much slower (i quit rendering after 7-8 minutes 'coz it still didn't rendered first pass).

Any explanation from NewTek Dev team / anyone ? This is really weird and it must be that something what is bothering LW9 in this particular scene 'coz most of scenes i've been doing are 2-3x faster with persp camera in Lw9 than in 8.5. But then agian it shouldn't be problem if 8.5 is rendering it fine it should be good in 9 but faster IMHO :).

jayroth
07-31-2006, 06:18 PM
It will likely have to wait til after Siggraph before we can take a good, hard look at this particular scene. It is an anomaly, though.

Jay Roth

Bytehawk
07-31-2006, 06:35 PM
My first thought was the fact that LW now has more accurate formulas for volumetrics, so I turned off trace shadows and set the light to shadow mapped

lw85 : 25 secs
lw9 : 51.3 secs

so.... even without tracing shadows for the volumetric light there is a 200% difference

jameswillmott
07-31-2006, 06:41 PM
I think I found it.

You have Opacity Cast Shadows turned on for your volumetric light. This never worked in 8.5, so now it has been fixed the scene takes longer because it's actually working correctly.

If you turn this function off you should get comparable times. 9 may still be slower, but it usually is (marginally) with simple scenes.

Musart
07-31-2006, 06:43 PM
Confirmed, LW 8.5 renders this scene in 57 secs (on my machine while surfing and doing some work what uses only 5-6% of CPU time - 3GHz P4) while LW9 (on both cameras - classic/perspective) is much, much slower (i quit rendering after 7-8 minutes 'coz it still didn't rendered first pass).

Any explanation from NewTek Dev team / anyone ? This is really weird and it must be that something what is bothering LW9 in this particular scene 'coz most of scenes i've been doing are 2-3x faster with persp camera in Lw9 than in 8.5. But then agian it shouldn't be problem if 8.5 is rendering it fine it should be good in 9 but faster IMHO :).

:agree:
It should at least the same or better... not worst

Lewis
07-31-2006, 06:43 PM
Good to hear Jay, thanks for quick response.

BTW Bytehawk what machine are you using ?

Musart
07-31-2006, 06:47 PM
I think I found it.

You have Opacity Cast Shadows turned on for your volumetric light. This never worked in 8.5, so now it has been fixed the scene takes longer because it's actually working correctly.

If you turn this function off you should get about a x2 speed INCREASE for 9 over 8.5...

:) That would be a cause ! It make sense

Musart
07-31-2006, 06:47 PM
My first thought was the fact that LW now has more accurate formulas for volumetrics, so I turned off trace shadows and set the light to shadow mapped

lw85 : 25 secs
lw9 : 51.3 secs

so.... even without tracing shadows for the volumetric light there is a 200% difference

I got 1 minute rendering with shadow map :)

Bytehawk
07-31-2006, 06:49 PM
I knew there was something changed over 8/5 in these things.

btw, i'm driving a :

Operating System: Win XP SP2
CPU Type: Xeon
Number of CPUs: 2
CPU Speed: 3 Ghz
RAM: 2 Gb
Graphics Card: nvidia fx 5900
Display Driver Version: 91.33
wacom intuos 3 - config set to tablet

Musart
07-31-2006, 06:54 PM
I think I found it.

You have Opacity Cast Shadows turned on for your volumetric light. This never worked in 8.5, so now it has been fixed the scene takes longer because it's actually working correctly.

If you turn this function off you should get comparable times. 9 may still be slower, but it usually is (marginally) with simple scenes.

Now i got : 1m 34s with opacity_cast_shadow - OFF :thumbsup:

Thanks

Is there anymore pseudo hidden not functioning thing like this elsewhere ??? :D

Lewis
07-31-2006, 06:57 PM
Hmm It is drasticaly faster when turning off "Opacity Cast Shadows" as James mentions but for me it's still slower than in Lw 8.5.

137.x sec with Classic and same time with perspective camera in Lw9?
57 sec in Lw 8.5 ?

jameswillmott
07-31-2006, 07:01 PM
LW's optimised for heavy polygon scenes, really simple ones may be slower than 8.5, but not by orders of magnitude :)

Lewis, I get much the same times as you do.

Volumetrics are far more correctly implemented in 9 than 8.5, which is what's slowing them down.

Bytehawk
07-31-2006, 07:04 PM
yep, still slower but not by that much anymore.
about same figures as you, Lewis.

again, turning on shadow maps in both versions with or without opacity casts volumetric shadows checked makes no difference: LW9 is still slower.

Lewis
07-31-2006, 07:10 PM
That's Ok James BUT if renders look same (i didn't notice any difference) who cares is it more mathematically correct when it's slower :)? It's not 2-3 secs slower it's slower 2x times in 9 than in 8.5 which is way too much in any case IMHO. Maybe Jay and Team still need to check what's under hood slowing it down so much.

Maybe we need another render engine called "Fake Camera" which is FAST and looking good but not rendering with deep/correct mathematics which will make it much faster in ALL things than any LW before :) :D.

J/K

jayroth
07-31-2006, 07:48 PM
Lewis,

James' explanations are correct, according to the render team lead. The feature never worked in 8.5, but now does in 9.

As for the two images appearing similar (with and without the bug), that may be true in this case, but not in most cases. Also, note that by turning off the feature that James mentions also produces the same result, so to make the assumption that 9 must be that much slower (and need some sort of correction) is invalid.

Hope this helps!

Jay

(and thanks, James!)

Musart
07-31-2006, 07:52 PM
Lewis,

James' explanations are correct, according to the render team lead. The feature never worked in 8.5, but now does in 9.

As for the two images appearing similar (with and without the bug), that may be true in this case, but not in most cases. Also, note that by turning off the feature that James mentions also produces the same result, so to make the assumption that 9 must be that much slower (and need some sort of correction) is invalid.

Hope this helps!

Jay

(and thanks, James!)

Well Thanks guys ... Now i understand what is happening. :thumbsup:

katsh
07-31-2006, 08:10 PM
my result was 44.4 secs.
my LW is 7.5d. it said some plug-in needed when open the scene.
but final image is correct i guess.

LW7.5d
Xeon 2.4Ghz x 2 (Dual CPU)
QuadroFX500 :newhere:

Musart
07-31-2006, 08:34 PM
my result was 44.4 secs.
my LW is 7.5d. it said some plug-in needed when open the scene.
but final image is correct i guess.

LW7.5d
Xeon 2.4Ghz x 2 (Dual CPU)
QuadroFX500 :newhere:

Yep the rendered image is correct

44 sec ... That's good !

Barred
07-31-2006, 10:46 PM
Musart I ran the scene just as you asked above. My rig is a P4 3.2 with 4 gig RAM and it took 1h 55m30s (6930.8s) in LW9. I will re-run it in LW8 tomorrow and post the time.

Weepul
07-31-2006, 11:15 PM
Mac user here.

LW 8.2: 17.7s
LW 9 after disabling Opacity Casts Shadows: 24.3s
8 threads in all cases.

The LW9 render looks different from renders from 8.2 both with Opacity Casts Shadows disabled and without: the volumetric effect is brighter in 9. The illumination on the object appears the same as in 8.2 with Opacity Casts Shadows disabled (but I do want to point out there was a significant decrease in illumination intensity on the object in 8.2 with Opacity Casts Shadows still enabled.)

Lewis
08-01-2006, 02:47 AM
Lewis,

James' explanations are correct, according to the render team lead. The feature never worked in 8.5, but now does in 9.

As for the two images appearing similar (with and without the bug), that may be true in this case, but not in most cases. Also, note that by turning off the feature that James mentions also produces the same result, so to make the assumption that 9 must be that much slower (and need some sort of correction) is invalid.

Hope this helps!

Jay

(and thanks, James!)

Thanks for effort Jay but mabye i didn't explain it well. I can understand that features are better now in 9.0 than 8.5 but my render times are with TURNED OFF that feature what james mentioned. I did turn off "Opacity Cast Shadows" for Vulumetric light and THEN LW9 renders 2x times slower than 8.5 (137 sec comparing to 58 sec in 8.5). If this feature never worked in 8.5 and since it's now turned off in 9 then it should be same image - right (atleast with classic camera)? When this feature is turned on (in LW9) then render times are horrid and it renders 60 times slower than in 8.5 (notice that users report 55minutes to 120minutes - not seconds). So with this feature OFF 9 is still much slower in this particular scene but what concernes me is that with this feature on is 60x times slower and i can't imagine what math calculation would be involved in such slowdown :). It's simply useles feature for any kind of animation and still images are also questionable if 640*480 resolution is rendered 55-120 minutes ? I can't imagine what would be time for 1600*1200 then (it would be rendered for day(s) ??)

I'm not saying Dev Render Team leader isn't right ('coz i know he is :)) but i'd still like that they check that render times after siggraph and that classic camera have same results like in 8.5 and I'm ok That perspective (new render engine) could have higher render times due more accurate rendering :).

Thanks

Musart
08-01-2006, 05:47 AM
Musart I ran the scene just as you asked above. My rig is a P4 3.2 with 4 gig RAM and it took 1h 55m30s (6930.8s) in LW9. I will re-run it in LW8 tomorrow and post the time.

Gee !!!

I'm eager to see your LW8 results

Mus

Chuck
08-01-2006, 09:37 AM
Thanks for effort Jay but mabye i didn't explain it well. I can understand that features are better now in 9.0 than 8.5 but my render times are with TURNED OFF that feature what james mentioned. I did turn off "Opacity Cast Shadows" for Vulumetric light and THEN LW9 renders 2x times slower than 8.5 (137 sec comparing to 58 sec in 8.5). If this feature never worked in 8.5 and since it's now turned off in 9 then it should be same image - right (atleast with classic camera)? When this feature is turned on (in LW9) then render times are horrid and it renders 60 times slower than in 8.5 (notice that users report 55minutes to 120minutes - not seconds). So with this feature OFF 9 is still much slower in this particular scene but what concernes me is that with this feature on is 60x times slower and i can't imagine what math calculation would be involved in such slowdown :). It's simply useles feature for any kind of animation and still images are also questionable if 640*480 resolution is rendered 55-120 minutes ? I can't imagine what would be time for 1600*1200 then (it would be rendered for day(s) ??)

I'm not saying Dev Render Team leader isn't right ('coz i know he is :)) but i'd still like that they check that render times after siggraph and that classic camera have same results like in 8.5 and I'm ok That perspective (new render engine) could have higher render times due more accurate rendering :).

Thanks

If the feature worked in 8.5, it would take as long or longer than 9.0 to render the scene with the feature turned on. Because the feature does not work in 8.5, 8.5 does not do the full set of calculations the feature requires. Because 8.5 does not do all of the required calculations, it renders the scene at a proportionally comparable speed to what 9.0 will render the scene when the feature is turned off in v9.

As to the difference in render speeds between the two when the feature is off, we've documented in our videos and and discussions of the changes to the renderer that for some scenes the older renderer may be faster. Smaller scenes particularly are most often faster, specifically due to the calculation overhead of the use of a KD-Tree algorithm to partition the scene for rendering versus the old method of partioning the scene for rendering. As scene complexity rises, though, the new method pretty quickly overtakes and surpasses the old method, and extends it's advantage the more you throw at it.

There are types of scene content where the older renderer may currently show an advantage in a more complex scene. In some cases this is because the new renderer is much less forgiving of things such as one and two point polygons (those need to be rendered with the Classic Camera), non-planar and degenerate polygons. In some cases we expect to be able to tune the KD-Tree algorithm and other aspects of the new, full-time ray-tracing renderer to account for these cases and bring the advantage to the new renderer versus the older scanline renderer with ray-tracing facilities.

Lightwolf
08-01-2006, 10:02 AM
It seems the node is using up some time as well. If you disable processing of the node, it should be up to LW 8.x speed again.

Nodes enabled: 1m19s (79.8s)
Nodes disabled: 49.5s

...and yes, it would be nice if they were disabled by default.

(btw, disabling "Use Texture" shaves off another 5 seconds here).

Cheers,
Mike

jayroth
08-01-2006, 10:17 AM
Folks,

In LightWave v9, you must make sure that your thread count matches the actual CPUs in your machine (hyperthreading counts, so you can count appropriately).

In the past, it was typical to allocate 8 threads. If you do so today, and you do not have 8 CPU nodes, you will not see optimal performance, as LightWave is now multi-threaded in an optimal manner.

So, if you have two CPUs, you should have two threads only. If you have two CPUs and each of them are hyperthreaded CPUs, then you can set your thread count to 4. If you have a single CPU, then you should set your thread count to 1.

If you overdo the thread counts, the rendering engine will be forced to perform unneccessary housekeeping, as well as run less efficiently. This will have a noticeable impact on your render times.

FYI,

Jay Roth

JamesCurtis
08-01-2006, 10:22 AM
Nice to know this!

Thanks Jay.

Lightwolf
08-01-2006, 11:06 AM
@jayroth
Is this true for the Classic Camera as well?
The new one re-allocate idle threads to the remaining segments, Classic doesn't seem to though.
Even then though, using more threads than cores rarely buys you a lot except if the scene is heavily unbalanced (in terms of rendering times per vertical segment).

Cheers,
Mike

colkai
08-02-2006, 02:45 AM
If you have a single CPU, then you should set your thread count to 1.

If you overdo the thread counts, the rendering engine will be forced to perform unneccessary housekeeping, as well as run less efficiently. This will have a noticeable impact on your render times.

Jay, just to show that computers always like to prove people wrong. :p

I've been doing several tests and though I've not tested it on RC23 yet due to time constraints, so far, I've found for my old single XP1800+ AMD CPU, I actually get improved performance at 2 threads rather than 1 on about 60% or more of my scenes.

Granted, I've not been doing exhuastive tests but the scenes I'm trying are pretty varied. Again though, yet to confirm this with RC23. Now I also need to check if this was PCAM or CCAM, though I'm pretty sure it was the former.

Barred
08-02-2006, 05:38 AM
Gee !!!

I'm eager to see your LW8 results

Mus


Mus I re-renderd the sceen in LW9 and had more or less the same time.
I rendered it out in LW8 and did exactly the same as I did in LW9 (load sceen and press F9) and got

56.2 sec

Something is wrong.

katsh
08-14-2006, 04:52 AM
my testing between v7.5 and v9.
some scenes , v9 is better speed.
some another scenes ,v7.5 is better speed.
i think,totaly,v9 is better than v7.5 about rendering speed.

Sensei
08-14-2006, 06:37 AM
It seems the node is using up some time as well. If you disable processing of the node, it should be up to LW 8.x speed again.

Nodes enabled: 1m19s (79.8s)
Nodes disabled: 49.5s

...and yes, it would be nice if they were disabled by default.

(btw, disabling "Use Texture" shaves off another 5 seconds here).

Cheers,
Mike

Hi Wolf!

That's strange.. By "disabling Nodes" you meant checkbox in Surface Editor, right?

On my Toshiba P100-219 laptop with Intel Centrino-Duo (dual-core CPU) 2.0 GHz, I got following results:

54.6 sec without nodes, 2 threads as Jay suggested
54.8 sec with nodes, 2 threads
51.4 sec without nodes, 8 threads
51.2 sec with nodes, 8 threads

Disabling "Use Texture" in the Volumetric Light Properties window saved 0.3 sec, too small to be really sure whether that's true time saved by this option..

With LightWave 8.5, same machine:

35.8 sec, 2 threads
31.2 sec, 8 threads

Best Regards!