OpenSubdivs

[FXDude]

Hi, Have any of you compared open subdiv performance with XSI Subdivs?

I assume they are probably somewhat faster, but by how much, and how are they generally?

At least XSI is has been left with the fastest (non-accellerated) subdiv system.
(regular subdivs are typically around twice as fast in XSI)

And still being the fastest of them all generally,(like changing things later, figuring out what's happenning, or making complex things using simple tools) OpenSubdivs or not.

But man.. I dont know how much of a feat it would be to implement, but they could have at least implemented currently available tech since they were currently available...

Anyways, Cheers!

[FXDude]

Hum.. Seems like Soft -still- has way faster subdivs...

http://forums.cgsociety.org/showthread.php?t=1227114
(pretty unanimous thread)

While also having (always have had) great interpolated UV inheritance.

[Mathaeus]

Well from what I read, OpenSubdiv is supposed to be faster in playback, should be able to generate displacement in viewport, so on. But only if base topology stays the same. Heard somewhere, it's actually slower in modeling, if there is any change in topology, like single extrude or add polygon (perhaps some heavy layers of data, just to get it to work). Also, every calculation that happens before, evaluation of rig an so on - it happens with unchanged speed. So yes, it seems to be tailored for Pixar's practice, not that much for anything else.
Pixar = almost unknown studio in game developing, most likely with zero experience when it comes to realistic characters, however they seems to be proud of their strange, artificial style of moving the puppets. Somehow interesting phenomena, some kids likes this, but all in all, almost irrelevant in real life....
As an personal point of view, entire subdiv thing was fascinating ten years ago, and SI in these times had about ten times faster subdivs than rest of the planet. Today, not that much of subdiv's glory - characters gets much more than 8 or 12 points for cross section of arm, so un subdivided display become more than enough for viewport. Slowly I'm trying to move the hard surface modeling to NURBS ( not Maya or SI), as I don't want to play with support lines for each detail to the end of my life.
Imho, what's called SI 'geometry engine', ability to do modeling, including topology changes, below envelope or another deformation op, without (almost) any penalty - *this* is still unmatched feature of SI. Others can do that, bit nobody can compete with SI, imho. I have a few characters who got their envelopes a years ago, and technically the same envelope remains, over almost completely changed topology ( including several re-creating of entire symmetry) through the years. But, someone first have to try this, to figure out what could be done - in other word, someone with habits formed by Maya, perhaps he won't even try to model below envelope, I think - that's the problem...

[Bullit]

OpenSubD's is less precise than Softimage modeling at least in beveling, the only appreciable advantage i remember is that if the implementations are the same it might work between DCC packages. So it can be edited and the booleans seems to be working better.

[luceric]

I think you're pretty on the money on things as always, Mathaeus, however one note about Pixar:
OSD is actually a collaboration with Microsoft, so it is definitely meant to apply to games. It's made to provide real-time playback flipbooks of fully tesselated and displaced render-quality characters. OSD is a work-in-progress - the first implementation they demoes was just for the DirectX11 viewport - and OpenCL and a few other features are not implemented yet in Maya 2015.

[luceric]

OpenSubD's is less precise than Softimage modeling at least in beveling, the only appreciable advantage i remember is that if the implementations are the same it might work between DCC packages. So it can be edited and the booleans seems to be working better.

I think you might be comparing the way Softimage implements its default normal calculation (the automatic discontinuity parameter) with other apps rather than something OSD does

[FXDude]

Previously, moving components on 2x subdivided ~100,000k triangle base mesh yeilded ~20FPS on XSI
and doing the same on the same mesh in Maya yeilded ~1 FPS with a few sec delay between each select->move operations,
so it seems that that didn't exactly change.

Yet if subd playback performance has been made more on par (by how much I dont know) then I guess that would make one more baby step ;)

[Bullit]

I mean the way Softimage bevel edges is more "real" and open than OpenSubD's.

[luceric]

Do you actually mean bevel, or edge creases?

[FXDude]

Yet if sub-d playback performance has been made more on par (by how much I dont know) then I guess that would make one more baby step ;)

Not that much more "on par" it seems, but better than nothing I guess...

From here

Might be worth starting a new thread for this, but I'm here now, so...

I grabbed one of our production rigs (animation rig) and compared playback frame rate with both smoothing methods, and with smoothLevel 1, 2 and 4. I'm getting better results with OpenSubdiv but only for the higher levels, and surprisingly only in the legacy viewport.

For legacy viewport:
Maya CC (level 1) 19 fps -- OpenSubdiv CC (level 1) 19 fps
Maya CC (level 2) 14 fps -- OpenSubdiv CC (level 2) 18 fps
Maya CC (level 4) 3 fps -- OpenSubdiv CC (level 4) 6 fps

For viewport 2:
Maya CC (level 1) 15 fps -- OpenSubdiv CC (level 1) 13 fps
Maya CC (level 2) 14 fps -- OpenSubdiv CC (level 2) 9 fps
Maya CC (level 4) 2 fps -- OpenSubdiv CC (level 4) 1 fps


This is on windows with a quadro 4000.

Probably not much can be concluded from this one test. Putting it here just for comparison. But on the face of it, these results certainly come as a disappointment given the hype surrounding both OpenSubdiv and Viewport 2 being the new default. New graphics card required maybe?

David
__________________

[Nox]

Ugh, that's a shame... considering all the fuss around OSD.

[Mathaeus]

By the way, according to words of Brad Peebler and Ton Rosendal, yeah it seems there is 'something' about OSD stuff. Links:
OSD discussion on Modo forums. - Brad's post at bottom of page.
Anticipated Blender projects of 2015

[Bullit]

Do you actually mean bevel, or edge creases?

I mean Bevel that is why i consider Softimage it superior. Creases for me is like cheating. Sloppy.

[FXDude]

Actually creases can be useful (for sharper corners without adding extra base geo)
and that's one thing that OSD does quite nicely by strategically allocating interpolated divisions towards the creased edge.

And I guess potentially having less base geo (when contol over slope is not necessary) is what can make it a bit faster than former Maya (not SI by a mile and a half) CtmlClrk subdivs.

But like with any subdiv system (including XSI), use of creases are quite a bit more prone to issues, and not as straight forward (visually or managability-wise) enough for many to just not use them.