Performance issues

[Herbert123]

Just a quick question regarding the sculpting performance in 3d-Coat. When I sculpt in Blender, and can go up to 50 million polygons (100 million triangles) sculpting an object. Now, orbiting becomes very laggy, but the "fast navigate" option solves that.

The point is, at that (granted ridiculous) resolution I can still sculpt for the most part smoothly with all tools, including masking. Going down to 20 million or less polygons gives me a smooth sculpting experience in Blender's sculpt mode.

Now, in 3D-Coat (DX) I have an object with about 20 million triangles (10 million polygons), and while orbiting is still good, sculpting is rather laggy and does not keep up with my strokes I make with the Wacom at all - it works smooth with a normal airbrush at a smaller brush size, but becomes extremely laggy with a textured brush. Especially slightly larger brushes with a texture start smooth for the first split second, but are very difficult to control when I keep dragging.

I am a bit confused why 3D-Coat is having performance difficulties on my system: i7 920@3.6GHZ, 48GB RAM, AMD 7970 3GB (to drive three screens), Nvidia GTX590 (for CUDA stuff only).

So, sculpting a 10 million polygon object is rather difficult in 3D-Coat on my system. I would expect me to be able to go much higher than that? Or are my expectations too high?

[AbnRanger]

Start with the splash screen in 3D Coat > Surface Mode sculpting > Base Human model (not Mannequin...the one to the right of it) > increase the res to 20mill> record your session. Export that same model to Blender and sculpt it with 50mill polygons and we'll see how big a difference there is. I'll record the same session with a 30mill poly version (3DC) and we'll compare notes...?

Funny, Michalis, a regular over on the Blender forums, and often drops by to do the same as you've done...bash 3D Coat and extol Blender in the same setting...says the poly limit is roughly 500k or so.

[Herbert123]

You misunderstand me - my intention was not to bash 3dCoat. I have been using the demo for some time now to get a feel for the program, and I think the sculpting is much more developed and more convenient than Blender's - I would compare it to ZBrush. I was just concerned that the performance seemed to be lacking.

Anyway, I figured it out: I was using an older demo version with CUDA disabled (non-cuda version?). I did what you suggested, and I can easily sculpt at 20million and more, and add details in voxel mode now.

And I realize now I am really comparing apples and pears. Both have completely different workflows: for example, in Blender sculpting works best with a limited base mesh, and upping it through the multi-res modifier. In 3dCoat no need for that. The layers are great. Upping the level to 92 million tris takes a long, long time, though. We are talking about 10 minutes. After that it works.But this is all theoretical - for real work I would never require that much resolution anyway.

In both apps it depends on the approach taken.

Thanks, I will keep testing and doodling.

[Carlosan]

Hi Herbert

Are you sculpting in VOXEL mode or SCULPT mode ?

[AbnRanger]

You misunderstand me - my intention was not to bash 3dCoat. I have been using the demo for some time now to get a feel for the program, and I think the sculpting is much more developed and more convenient than Blender's - I would compare it to ZBrush. I was just concerned that the performance seemed to be lacking.

Anyway, I figured it out: I was using an older demo version with CUDA disabled (non-cuda version?). I did what you suggested, and I can easily sculpt at 20million and more, and add details in voxel mode now.

And I realize now I am really comparing apples and pears. Both have completely different workflows: for example, in Blender sculpting works best with a limited base mesh, and upping it through the multi-res modifier. In 3dCoat no need for that. The layers are great. Upping the level to 92 million tris takes a long, long time, though. We are talking about 10 minutes. After that it works.But this is all theoretical - for real work I would never require that much resolution anyway.

In both apps it depends on the approach taken.

Thanks, I will keep testing and doodling.

Yeah...I was testing 30mill just now in Surface mode and it's ridiculously fast. Brushing anyway. Using the MOVE tool, there seemed to be some initial lag still, but I'm running on a backup PC (my main one went down this past week) with RAM running at 1066 (RAM module's rated to 2000mhz, but the Intel board is so finicky I can only run it at base setting), and an i7 950. I find that fast memory speeds makes a BIG difference...maybe more so than using CUDA. Even at slow speeds though, I'm pretty amazed at how much faster the sculpting seems to be lately. Andrew said there was a 3-10x increase, after chasing down some lag issues in the Retopo room.

I usually try to cache the layer anyway, before I try to use the MOVE or POSE tool on a dense mesh. At 30mill, it's so dense, you can't even see the wireframe with it toggled on...unless you zoom right up on the model.

[BeatKitano]

Which is nice until you try to make nice crisp (not jaggy) edges, and then the dyntesselation falls appart.

Performances are important, quality/control of strokes is too. I'd even venture as far as saying that the speed would get quite a hit if the surface sampling was done right...

[AbnRanger]

If I adjust my brush settings accordingly, CreaseClay usually treats me well, as does the standard Pinch tool. So, I'm not seeing this as an issue yet. I do notice some individual spots/blobs sometimes, that I plan to ask Andrew about. But I can usually get rid of them after adjusting some of the brush settings...even the depth profile. I noticed today, when I went to adjust it at one point...on one particular brush/tool the depth profile was wacked out....all crazy looking. I reset it to a slight S-Curve and it was good to go.

I do think there is definite room for improvement, no doubt. I also noticed some holes created at one point, and when I used CLOSE HOLES on them, it locked my PC up. I waited and waited several minutes and ended up having to reboot. I sure hope Raul is able to return or get a good internet connection there in CUBA. I can only imagine how much better things could be if he was still on staff with them.

[artman]

Which is nice until you try to make nice crisp (not jaggy) edges, and then the dyntesselation falls appart.

Performances are important, quality/control of strokes is too. I'd even venture as far as saying that the speed would get quite a hit if the surface sampling was done right...

sorry,but no jaggy edges on my side....actually i get less clean strokes using Damstandard (unless I subdivide mesh to a ridiculous level).

[digman]

Not a problem getting clean edges on my end either... CreaseClay and surface pinch with your own alphas and settings will do the job quite well as AbnRanger stated... You got the new spline tool that is very handy also...

Tsmooth (tangent turned off) plus using the relaxation setting will relax the mesh and smooth right up to the edge without ruining them.

[ajz3d]

Start with the splash screen in 3D Coat > Surface Mode sculpting > Base Human model (not Mannequin...the one to the right of it) > increase the res to 20mill> record your session. Export that same model to Blender and sculpt it with 50mill polygons and we'll see how big a difference there is. I'll record the same session with a 30mill poly version (3DC) and we'll compare notes...?

Funny, Michalis, a regular over on the Blender forums, and often drops by to do the same as you've done...bash 3D Coat and extol Blender in the same setting...says the poly limit is roughly 500k or so.

On my workstation, in the scene with the human model at 26mln triangles, it's:

- impossible to sculpt anything in the voxel mode with Draw brush set to radius of four and CUDA smooth boost disabled.

- fairly OK to sculpt with the same radius and CUDA smooth boost enabled (but with visible "dottiness").

- In surface mode, it's a little bit better, but I wouldn't call it ridiculously fast.

System specs: 3930K, 32GB RAM @1600MHz CL10 quad channel, GTX660Ti 3GB.

Can you please post a video on sculpting at 26mln tris (2x Res+ on the human model) in voxel and surface modes? I'm getting a bit worried about performance of my system here.

Edited June 2, 2013 by ajz3d

[BeatKitano]

I really have to make a comparative video with zbrush to show what I mean. Can't right now (broken ankle and torn muscles) but definitely will. Cause I can't believe I'm the only one seing a big flaw here.

Everytime I bring this to the table someone says "play with the settings", maybe it works that way but my point is with zbrush (and others) you don't need to tweak every setting to get clean, well defined edges.

Currently the only way to get something close to zbrush is to use pinch brush with a large strength value. That's ok, except you have to use a large radius to get the desired effet and then you screw up the surround surface. AND NOW you have to work masks.

See: zbrush (and others) simple brush association for a task;3dcoat : tweaking brushes, using masks and alternative smoothing technics...

I don't know maybe it's just me but I get cleaner edges (even if the topology doesn't match the shape) by using dynamesh and reproject, than using brushes and alll those tricks with liveclay in 3dc.

NOT EFFICIENT. But then again it's not like I said it already ▼

[AbnRanger]

I really have to make a comparative video with zbrush to show what I mean. Can't right now (broken ankle and torn muscles) but definitely will. Cause I can't believe I'm the only one seing a big flaw here.

Everytime I bring this to the table someone says "play with the settings", maybe it works that way but my point is with zbrush (and others) you don't need to tweak every setting to get clean, well defined edges.

Currently the only way to get something close to zbrush is to use pinch brush with a large strength value. That's ok, except you have to use a large radius to get the desired effet and then you screw up the surround surface. AND NOW you have to work masks.

See: zbrush (and others) simple brush association for a task;3dcoat : tweaking brushes, using masks and alternative smoothing technics...

I don't know maybe it's just me but I get cleaner edges (even if the topology doesn't match the shape) by using dynamesh and reproject, than using brushes and alll those tricks with liveclay in 3dc.

NOT EFFICIENT. But then again it's not like I said it already ▼

Using the right tool for the job isn't a trick. CreaseClay isn't a trick. It's a tool. I think there is lots of room for improvement, too, though, because I still get these random dollops/blobs, and that forces me to stop and UNDO or waste extra time having to smooth those out. Once you get in a groove, you should have to keep getting interrupted. I don't know if its the Wacom Intuos 3 drivers on Win 8 being wonky, or if it's something Andrew can correct.

[BeatKitano]

Using the right tool for the job isn't a trick. CreaseClay isn't a trick. It's a tool. I think there is lots of room for improvement, too, though, because I still get these random dollops/blobs, and that forces me to stop and UNDO or waste extra time having to smooth those out. Once you get in a groove, you should have to keep getting interrupted. I don't know if its the Wacom Intuos 3 drivers on Win 8 being wonky, or if it's something Andrew can correct.

My apology then, it seems you know what I mean

Getting the same thing done in half the time in another app is clearly the symptom of a real issue. I'm certainly not the most efficient guy with 3dcoat since I always used it as I would with real clay (background I suppose), but even though 3dc definitely got the better "numerical equivalence" in that regard d it clearly is behind regarding brush/workflow efficiency.

I don't say it's crap (if it was the case I would not have been there since 2008) but it's clearly missing the spot a few area (sculpting wise, because like I said earlier that's my game).

[AbnRanger]

On my workstation, in the scene with the human model at 26mln triangles, it's:

- impossible to sculpt anything in the voxel mode with Draw brush set to radius of four and CUDA smooth boost disabled.

- fairly OK to sculpt with the same radius and CUDA smooth boost enabled (but with visible "dottiness").

- In surface mode, it's a little bit better, but I wouldn't call it ridiculously fast.

System specs: 3930K, 32GB RAM @1600MHz CL10 quad channel, GTX660Ti 3GB.

Can you please post a video on sculpting at 26mln tris (2x Res+ on the human model) in voxel and surface modes? I'm getting a bit worried about performance of my system here.

First of all, notice I said "Start from the splash screen and choose SURFACE mode Sculpting" At that level density, It's not very practical or smart to be sculpting out rough forms, but rather high detail work....which it does fine, even in Voxel mode. But it is noticeably faster, even...yes, ridiculously fast...in Surface mode. I tried to keep the doddle short, but after a few attempts, I gave up. It's about 17min, if you want to see still. I even showed how different brush alphas and draw modes can greatly affect the ability to get a clean, hard edge when using the Pinch tool (used it in Voxel mode...would have been even sharper in Surface mode). A gradient draw mode and soft gradient brush alpha makes it work pretty well.

http://www.screencast.com/t/jwyHgxMSdZKF

With all of that said, it's been mentioned over and over on this forum....use Voxel mode for your early stages of sculpting...even up to a Medium - Med-High resolution (which in this case 10mill would have been the limit I would have gone to before switching to Surface mode). You also have to keep in mind that your card and mine are 600 series Kepler cards, and the CUDA performance in these cards are a step back from the Fermi cards. They navigate about the scene much faster, but CUDA performance isn't quite as good. Also, on that 26mill model, toggle on wireframe and navigate about the scene. See how slow and choppy it is? This 600 series is a pretty big disappointment for me already (mine is a 670 4GB)...and Andrew never has recompiled CUDA (it was coded back when CUDA was still on v1...we're at 5.5 now and lots of new tech has been added but not access by 3D Coat because Andrew doesn't want to waste time recompiling it....thinking it won't make that much difference. However, that was his same thinking about Multi-threading. Once he finally did it, it made all the difference in the world.

http://3d-coat.com/mantis/view.php?id=772

[Carlosan]

Thats the point

To use only the tools isnt enough to achieve good results

3DC is very powerful about tool parameters but

is WEAK in PRESETS management

I think Andrew MUST:

- focus in the PRESET AREA,

- give a default set of presets in every release

- and create a NEW subsystem to administrate its in a very clear way for the new and old users.

[BeatKitano]

Yeah but tweaking the existing parameters won't help in some cases.

3dc got the worse sampling system in existence. That, to me is a fact.

It has spacing, direction (view/vert normal) and that's about it. Just go into stroke and brush>samples/modifiers see how all those parameters are used in a way or another to get creative brushes like the polish/tirm dynamic brushes and such.

in 3dc you don't have that. The sampling is very weak.

Try this: create a ridge following the underlaying surface. Try to flatten that ridge to make it "more solid", try to achieve that "mechanical" look at the base (crease and flatten at the edges). You'll have a hard time getting this "perfect" (hard surf perfect is a dream ^^) and will spend most of your time redefining your edges/smoothing out imperfection (jaggy edges).

In zbrush you can pretty much to that in half the time with a few settings (preserve edges in brush>samples for instance).Andrew tried to open a bit the parameters (the infamous general brush was born from this very idea, unfortunately I don't think anyone asked for ONE brush doing everything with parameters fighting each other.)

[AbnRanger]

My apology then, it seems you know what I mean

Getting the same thing done in half the time in another app is clearly the symptom of a real issue. I'm certainly not the most efficient guy with 3dcoat since I always used it as I would with real clay (background I suppose), but even though 3dc definitely got the better "numerical equivalence" in that regard d it clearly is behind regarding brush/workflow efficiency.

I don't say it's crap (if it was the case I would not have been there since 2008) but it's clearly missing the spot a few area (sculpting wise, because like I said earlier that's my game).

If you can record a setting in 3DC to illustrate your issue and then perform/record the same thing in ZBrush and email that to Andrew or post it here, Hopefully he can do something to rectify it. A lot of features I've requested, I sent him a recording to illustrate just why it was needed. That gets his attention more than anything.

For example, I was able to share the screen with him to show how 3D Coat was lacking compared to Mudbox when working with large brush sizes and large maps. He kind of dismissed it and said to just use smaller brushes. That's not good. Suggesting we settle for only using small brushes is not a solution. I then showed him an example of why that just isn't practical, many times. For one, if you are trying to apply a stroke (with spacing) with a certain brush alpha (ie, Reptile scales), using a smaller brush won't work when you are trying to sculpt or paint large scales/patterns. This issue MUST get fixed...that's why I keep harping on it.

[BeatKitano]

If you can record a setting in 3DC to illustrate your issue and then perform/record the same thing in ZBrush and email that to Andrew or post it here, Hopefully he can do something to rectify it. A lot of features I've requested, I sent him a recording to illustrate just why it was needed. That gets his attention more than anything.

Well, I think I got a few of them in the voxel room in my sig didn't help much :/

Anyway if this was my call, I would just create a "blank slate" version of the app, with all parameter exposed, add new one if a behaviour is requested, and let users create brushes matching thir taste. No doubt that if a brush is good enough it'll be adopted in no time and can be defined as "standard"... Dam Standard is a pretty good example in zbrush, it was liked so much they included it in zbrush in standard now.

I know this pretty much sounds like the preset system, except the preset system is a "try to emulate" tool. Not a painter's palette.

[AbnRanger]

Well, I think I got a few of them in the voxel room in my sig didn't help much :/

Anyway if this was my call, I would just create a "blank slate" version of the app, with all parameter exposed, add new one if a behaviour is requested, and let users create brushes matching thir taste. No doubt that if a brush is good enough it'll be adopted in no time and can be defined as "standard"... Dam Standard is a pretty good example in zbrush, it was liked so much they included it in zbrush in standard now.

I know this pretty much sounds like the preset system, except the preset system is a "try to emulate" tool. Not a painter's palette.

I can tell you one thing for sure...if you present Andrew with a massive post or e-mail there is a 100% chance he won't read it. It can be very time consuming and he probably gets lots of those. So, keep it brief and record it on a video so he doesn't have to spend all day sorting through a 3page essay....cause he won't. I have sent some with a few paragraphs and never get a reply. A few sentences and he usually will try to respond.

[BeatKitano]

Said topic was a first draft with what... 20% of what needs changes... I even stated it, he read the thing (supposedly since he responded to it).... and nothing.

I'm at this point where I still hope that somehow Andrew take notice of those issues, but at the same time don't really believe in it since this thread is the third of the kind, is from last year, and nothing has changed since.

I'm getting tired to repeat myself to only get echo of my own voice (I suppose you share this sentiment with the cuda issue). This is also why I recently started to disseminate a few bits of info regarding HeLix... the thing is not finished but at least every bit of feedback from beta users is taken into accound and somehow integrated (sometimes after a huge change in the way the app handle things) still hoping Andrew takes notice.

It's cohesive, streamlined and it give users complete access to key elements.

Something that has been missing since the early days of 3.x (and the addition of preset system is only a partial answer).

Oh and AbnRanger: ADHD right ? :x

[ajz3d]

First of all, notice I said "Start from the splash screen and choose SURFACE mode Sculpting" At that level density, It's not very practical or smart to be sculpting out rough forms, but rather high detail work....which it does fine, even in Voxel mode. But it is noticeably faster, even...yes, ridiculously fast...in Surface mode. I tried to keep the doddle short, but after a few attempts, I gave up. It's about 17min, if you want to see still. I even showed how different brush alphas and draw modes can greatly affect the ability to get a clean, hard edge when using the Pinch tool (used it in Voxel mode...would have been even sharper in Surface mode). A gradient draw mode and soft gradient brush alpha makes it work pretty well.

http://www.screencast.com/t/jwyHgxMSdZKF

With all of that said, it's been mentioned over and over on this forum....use Voxel mode for your early stages of sculpting...even up to a Medium - Med-High resolution (which in this case 10mill would have been the limit I would have gone to before switching to Surface mode). You also have to keep in mind that your card and mine are 600 series Kepler cards, and the CUDA performance in these cards are a step back from the Fermi cards. They navigate about the scene much faster, but CUDA performance isn't quite as good. Also, on that 26mill model, toggle on wireframe and navigate about the scene. See how slow and choppy it is? This 600 series is a pretty big disappointment for me already (mine is a 670 4GB)...and Andrew never has recompiled CUDA (it was coded back when CUDA was still on v1...we're at 5.5 now and lots of new tech has been added but not access by 3D Coat because Andrew doesn't want to waste time recompiling it....thinking it won't make that much difference. However, that was his same thinking about Multi-threading. Once he finally did it, it made all the difference in the world.

http://3d-coat.com/mantis/view.php?id=772

Ya, my bad, sorry. I didn't read your previous post well. I fired up 3D Coat just to do some quick doodles with giant brushes over high density voxel volume, without too much thinking. Bad idea, like you already pointed out. Guess I forgot the principles.

When starting from the surface mode the 20mln poly sculpting does indeed feel exceptionally fast with medium sized brushes.

The wireframe view performance indeed is terrible on 600s, but it doesn't bother me that much because I rarely use it in the sculpt room, though I really keep my fingers crossed for the Mantis report you linked up and I strongly support it. Maybe one day Andrew will realise that it's worth to recompile the code to support CUDA 5.5+.

[AbnRanger]

My AMD (X6 1090t) board went out this past week and I'm using a backup (i7 950). I was finally able to get this finicky Intel Board to bump the RAM speed up to 1400mhz (doesn't like the 2000mhz RAM I put it in it), but stock CPU speed...and even with that, I did a little test in the Paint Room. Where I noticed Large brush lag on 4k maps or bigger, I was gobsmacked to find that it's not bad at all on a 4k map. So, it seems 3D Coat really prefers Intel CPU's. This stock i7 950 (3.3Ghz with Turbo Core) seems faster than the AMD X6, which was running at 4.1GHz.

I recall some talk back around Siggraph 2010 that Andrew had purchased some Intel libraries that would help boost performance to some degree. So, maybe that explains it. Both CPU's were roughly neck and neck in most of the benchmarks I saw, so I went with the cheaper of the two...plus the AMD boards are always less expensive.

Getting back to CUDA...when Andrew added CUDA Smooth Boost a while back, it made a big difference...otherwise I'd have to cache the layer and smooth the proxy. That's fine sometimes but CUDA Smoothboost is ready and waiting to be used instantly. No waiting...unless you are in Surface mode. I really wish Andrew would hire a GPU programmer to either recompile and expand CUDA's usage or switch to OpenCL and have a trigger point of a certain number of verts,voxels, or pixels, before the GPU kicks in and does the Parallel threading.