Distant objects, 2 passes, 2 far planes, clearing zbuf, zbuf depth

We're using 2 pass in many of our projects (mobile devices). It works fine. But you will lose some benefit from hw provided hidden surface removal function, and also other function depending on z buffer (texture). For the far-pass, we have less polygon and light pshader than near pass, and almost of things in far-pass are solid. And we didn't clear zbuffer before the 2nd pass, just set the depth range.

Before you try something that drastic, try experimenting with your near plane. The far plane has very little impact on z-fighting (counter-intuitively, sometimes increasing the far plane can help resolve z-fighting issues!), whereas the near plane has a huge impact.

This is because the hardware depth buffer is hyperbolic. As a very rough rule of thumb, imagine half of your precision going towards representing the objects between near and 2*near, and the other half representing objects from 2*near to far.

[edit]You can use this form to see when 24 depth buffers were introduced:

http://zp.amsnet.pl/cdragan/query.php?dxversion=9&feature=formats&featuregroup=selected&adaptergroup=groups&adapterselected%5B%5D=ATi&adapterselected%5B%5D=Intel&adapterselected%5B%5D=NVidia&featureselected%5B%5D=39&featureselected%5B%5D=41&featureselected%5B%5D=42&featureselected%5B%5D=44&featureselected%5B%5D=46&resource=SURFACE&usage=DEPTHSTENCIL&orientation=horizontal

Looks like you'll be fine, except for ancient Intel cards. If you're going to support these, I'd recommend getting one off ebay for a few bucks so you can test the framerate on it too

[edit2] That list is for the automatic depth buffer that you get with your device.

If you want to create extra depth buffers, or depth buffers that can be read as textures, then support is very spotty until DX10-era cards. e.g. there's some cards that support an automatic 24-bit depth buffer, but only 16-bit depth buffers if you want to be able to bind them as a texture.

And we didn't clear zbuffer before the 2nd pass, just set the depth range.

clever. guess you had the depth to spare.

in the title i mentioned, a 16 bit zbuf is only good out to a range of about 300 or so, given the meshes being drawn, and how close their faces can be in Z. 16 bit zbuf and 1000 far plane causes some z fighting starting at about 250 range.

in another title i'm working on, a flight simulator, visible range is 10000 d3d units at 100 feet per d3d unit (189.39 miles). and the largest target (an airship) is visible out to 8000 d3d units (151.51 miles).

This is because the hardware depth buffer is hyperbolic. As a very rough rule of thumb, imagine half of your precision going towards representing the objects between near and 2*near, and the other half representing objects from 2*near to far.

non-Linear?! well that explains a lot!

ok, i have the near plane at 0.1 d3d unit. in that title, the scale is 1 foot = 1 d3d unit. .0.1 is required to fit the player inside a small lean-to or shelter (pup-tent sized).

so, well i was going to say i should use a 0.1 near plane only for location=INSHELTER, but you can still see the world, and z fighting at 250 range and 16 bit zbuf...

since shelters are a world object, i can't create a "bigger" scene or model that a bigger near plane will fit in. think of a model of a pup-tent, with one end open. now put the camera inside the pup tent in the center, about 1 foot off the ground. the tent is about 3 feet tall. its pretty tight in there. but a nice effect, like you're stretched out on your sleeping bag, gazing out at the world.

i did notice that z fighting got a little worse when i moved the near plane from 1 to 0.1 to squeeze it inside a shelter.

also i do recall something about using as far a near plane as possible, during my last round of directx optimization research.

sounds like i can rely on 24bit being available, and i can do two (or more) passes if needed.

what about setting the near plane of the 1st pass to the far plane of the second, or a little inside it:

pass 1: near=450 far=10000

pass 2: near = 0.1 or 1, far = 500

that would keep the zbuf resolution of pass 1 higher than near=1 far=10000.

that way you could "stack up" zbufs one behind the other, draw out to any range with multiple passes, and still get a high resolution zbuf for each pass.

i think that may be the answer.

fortunately, it looks like i can count on a 24bit zbuf being available for the primary swap chain, so hopefully I won't have to go to a 2 pass method.

And we didn't clear zbuffer before the 2nd pass, just set the depth range.

clever. guess you had the depth to spare.

in the title i mentioned, a 16 bit zbuf is only good out to a range of about 300 or so, given the meshes being drawn, and how close their faces can be in Z. 16 bit zbuf and 1000 far plane causes some z fighting starting at about 250 range.

in another title i'm working on, a flight simulator, visible range is 10000 d3d units at 100 feet per d3d unit (189.39 miles). and the largest target (an airship) is visible out to 8000 d3d units (151.51 miles).

Plz check this page, http://www.sjbaker.org/steve/omniv/love_your_z_buffer.html

near value is more important than far value.

1. use 16 bit zbuf for compatibility
2. set far plane to high value (say 10,000)
3. draw distant stuff
4. set far plane to low value (300 or so)
5. clear zbuf
6. draw close stuff

Personally I wouldn't touch 16bit z with a ten-foot pole. Standard is 24bit+8bit stencil. Even 1st gen Atom systems support it as far as I recall and I'd advice you to not even try using anything lower in perf.

By the way, I I experimented the "double frusta" method you mention, albeit I couldn't look it in depth for example the performance implication with shadows and such.

In short: it works... sort of. But I couldn't manage to fix a few missing pixels on the "zMiddle" plane. In certain cases, those were quite noticeable.

i implemented a Zset_clip_planes(near,far) routine.

now i set the clip planes to 400 and 2000, and draw clouds. then i set the clip planes to 0.1 and 1000, and draw everything else.

works great.

Nice to know. No missing pixels?

Nice to know. No missing pixels?

By the way how many times can you do this trick? I mean, does it work for really large distances or do you eventually lose all precision for very high near and far values (say you were rendering very distant objects). Or do you use a different method for this?

I have experimented briefly with this. I did a far pass with near=100,far=10,000 then cleared the depth buffer and did near=0.1,far=100. It looked great except for some popping around the border, so I tweaked it so the two ranges overlapped slightly. Problem solved. I'm not sure how the framerate works out.

Nice to know. No missing pixels?

no problem.

but then again, the way i'm doing things, i don't think there could be.

for daytime scenes, i:
1. clear the screen to sky color (for the moment).
2. clear the zbuffer.
3. set lights based on time of day.
4. set camera
6. beginscene
7. draw sun: ztest off. zwrite off. lighting off. alphablend on. one drawprimitive call, one static quad mesh (2 tris), one static texture. then turn lighting, zwrite,and ztest back on, and alphablend off again.
8. draw clouds: set material based on time of day. set clip planes to 400 and 2000. recalc frustum clipping planes. add 1000 clouds to the "drawlist" - a list of textures in the scene, and the meshes that use them. draw the meshes in the drawlist: for each texture, set texture. for each mesh that uses that texture, set world transform. set mesh, material, cull, alphatest, and clamp on the fly as needed using a state manager. then call drawprimitive. reset clip planes to 0.1 and 1000. recalc frustum culling planes.
9. clear zbuf.
10. draw everything else
11. present

so i clear the screen and the zbuffer, then alpha blend in the sun with zbuffer test and write turned off. then i turn on the zbuffer again and draw the clouds with the distant clip planes. then i clear the zbuffer and draw everything else with the closer clip planes.

i may not need to clear the zbuffer between drawing clouds and everything else. the near plane for clouds is 400. visual range (for non LOD background stuff) is only 300. so i never really draw anything much past 300. and when i do clear it, the zbuffer just has ~500 clouds starting at range 400, which won't change anything if i then draw a bunch of stuff at 300 or closer.

By the way how many times can you do this trick? I mean, does it work for really large distances or do you eventually lose all precision for very high near and far values (say you were rendering very distant objects). Or do you use a different method for this?

Zset_clip_planes() calls directx's set projection matrix routine. so that's the overhead hit you take every time you change the clip planes.

i've been playing with very large clip planes in my airship flight sim. visual range there for the largest object (a size XL zep) is almost 200 miles. so its sort of like doing a space flight sim, a few of which i've done in my day. in the airship flight sim, the scale is 1 d3d unit = 1 foot. the clip planes are set to 10 and 1000000 (1 million). the "world map" is 9030 miles across (from the Ural Mountains in Russia to about Chicago in the United States). needless to say, d3d's coordinate system can't handle that size of a world at that scale. so the world is divided up into 10000x10000 sectors, each with its own local d3d coordinate system.

i see no reason why the far clip plane can't be set out to the limits of the d3d coordinate system's ability to draw things. and the call to set projection matrix is the only cost. so the thing to do would be draw the scene in sections from most distant to closest. for each section, you'd set the planes to just enclose that section. this would give you maximum zbuf accuracy for the upcoming drawcall, to avoid zfighting. then you set the planes to just enclose the next section closer, and draw that. then you set the planes for up close and draw the rest. then you might want to set them very close to draw some high rez mesh up close, such as a weapon in hand.

sections might be split into sky (background) stuff, distant lod (3d skybox) stuff, normal visual range stuff, and closeup stuff. perhaps even 2 distant lod or normal range sections, depending on the total visual range and scene complexity. by using this "divide and conquer" strategy combined with good old painter's algo of farthest to nearest, one should be able to render scenes of arbitrary visual range with no zfighting.

actually, set projection matrix is not the only cost. you have to clear the zbuffer for each section draw, for max zbuffer accuracy. but you may have sufficent zbuffer accuracy to span 2 or more adjacent sections between zbuffer clears.

come to think of it, you'd have to clear the zbuf every time you reset planes, as the zbuf values are a function of the clip planes since its non-linear. at least i suspect so.

so zbuf data would only be valid for a given set of clip planes. you could for example draw some stuff, then turn off zbuf , change clip planes, draw more stuff, then change back to the old clip planes, and turn zbuf back on, and your zbuf data would still be valid for further drawing. but zbuf use with different clip planes would require a zbuf clear, i'm pretty sure. can anyone out there confirm this?