Jump to content

  • Log In with Google      Sign In   
  • Create Account

Depth Buffer and deferred rendering


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
13 replies to this topic

#1 pancakedice   Members   -  Reputation: 159

Like
0Likes
Like

Posted 28 July 2010 - 05:24 AM

Hi,

I'm currently working on a deferred renderer and now trying to optimize the lighting calculation. The way I have done it so far is the a color attachment to store depth value linearly, which I then use to reconstruct position. Then, I figured that I need to do depth testing for more efficient lighting calculation - or rather not to do the calculations. However, early z does only work with the default depth buffer as far as I understand, and thus I have to duplicate depth information, stored in two different ways.

How do you normally solve this? Use the default depth buffer and letting opengl write to that itself (possibly flipping far and near for more precision), or are you simply saving the info in a dedicated color attachment? If you save it in a depth attachment, do you do it as a texture or as a renderbuffer? Can you do depth testing on a MRT with an attached texture as depth attachment, and is there a performance penalty for this?

(The way I currently store depth is the negated view-space z value divided by zfar, and I figure I need to fetch it much differently if I get it as z/w after perspective transformation from the standard depth buffer)

Thanks in advance.

Tell me if I need to elaborate anything.

Sponsor:

#2 allingm   Members   -  Reputation: 452

Like
0Likes
Like

Posted 28 July 2010 - 09:32 AM

If I understand you correctly you want the precision of linear depth and are worried about using two depth buffers. You have no choice. You simply have to choose between using the hyperbolic depth alone, or using both the hyperbolic depth and linear depth. Do you need the extra precision? If so, use two textures. If you don't then just reconstruct your position from the hyperbolic depth. I'm assuming you are using OpenGL which should allow you access to reading the depth buffer. DirectX 9 and below do not allow you to do this. You can find more information from the articles below.

You can find more information about the depth buffere and precision here:
http://mynameismjp.wordpress.com/2010/03/22/attack-of-the-depth-buffer/

You can find more information about reconstructing position from depth here:
http://mynameismjp.wordpress.com/2009/03/10/reconstructing-position-from-depth/

#3 pancakedice   Members   -  Reputation: 159

Like
0Likes
Like

Posted 28 July 2010 - 09:56 AM

Thanks - hyperbolic depth might have been the keyword I was looking for when describing the situation.

I have read those articles, but thanks for the link anyway. If I deciphered your message correctly, then you recommend that I just use the hyperbolic depth buffer, and then reconstruct the position from that, given it provides me with enough precision.

Then, for the second part. Yes, I am using OpenGL. For reading back from the depth buffer, can you read the depthbuffer directly or do I need to attach a texture to a depth attachment, ie:

glBindTexture(GL_TEXTURE_2D, depth);
glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
glTexImage2D(GL_TEXTURE_2D, 0, GL_DEPTH_COMPONENT, width, height, 0, GL_DEPTH_COMPONENT, GL_UNSIGNED_BYTE, NULL);
glFramebufferTexture2D(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_TEXTURE_2D, depth, 0);



Or is it possible to read from the "default" depth buffer where I do not add a depth attachment? (glReadPixels is slow, so I guess another way ;) )

What if I use a renderbuffer as a depth attachment, i.e.:

glGenRenderbuffers(1, &fbo.depth);
glBindRenderbuffer(GL_RENDERBUFFER, fbo.depth);
glRenderbufferStorage(GL_RENDERBUFFER, GL_DEPTH_COMPONENT24, fbo.width, fbo.height);
glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, fbo.depth);


I'm not quite sure what the driver does internally. Does it always reserve RAM for the "standard" depth buffer, and does adding my own as either a texture or renderbuffer have any cons?



#4 venzon   Members   -  Reputation: 256

Like
0Likes
Like

Posted 28 July 2010 - 10:27 AM

I use a perspective z/w depth buffer (the native opengl depth buffer format) of 24-bits for my deferred renderer and am able to reconstruct position with no problem.

To be able to read the depth buffer you should: 1) use a depth format texture attached to the depth attachment of your FBO on your first pass 2) bind the depth format texture on your second pass. If you do "glTexParameteri(texture_target, GL_DEPTH_TEXTURE_MODE, GL_LUMINANCE);" you can (inside your shader) just sample from the depth texture like any other texture, and you'll get the full-precision floating point value back in (for example) the .r swizzle.

#5 pancakedice   Members   -  Reputation: 159

Like
0Likes
Like

Posted 28 July 2010 - 11:22 AM

Thanks.

If I write depth info to a texture attached as a depth attachment in my geometry pass, how will I then be able to reuse that particular texture as depth buffer for z-fail testing afterwards? I know how to sample it, but would like the glEnable(GL_DEPTH_TEST) to work. :)

Thanks so far.



#6 pancakedice   Members   -  Reputation: 159

Like
0Likes
Like

Posted 29 July 2010 - 08:29 AM

So far I am rendering to a depthbuffer attached as a GL_TEXTURE_RECTANGLE_ARB, and then for the lighting pass retrieving the information as non-linear depth, converting it and everything is fine. However, I would like to know how I can reuse the depth texture for standard glDepthTest afterwards, when I have unbound the framebuffer object in order to draw to the window framebuffer.

Any help is appreciated.

#7 My_Mind_Is_Going   Members   -  Reputation: 132

Like
0Likes
Like

Posted 29 July 2010 - 03:45 PM

If you add a depth attachment to the render target where you do your final lighting pass, you can copy the depth buffer from your gbuffer over to it with glBlitFramebuffer and then enable GL_DEPTH_TEST like you want. As far as I can tell though you can't use glBlitFramebuffer to blit depth to the back buffer, so I do it to an intermediate render target and then blit the color from that to the screen at the end.

#8 allingm   Members   -  Reputation: 452

Like
0Likes
Like

Posted 29 July 2010 - 08:40 PM

You wanted 1 depth buffer right? I'll assume so in the following.

Are you trying to write to a depth buffer manually and then get the hardware to do depth testing with it? That seems a little odd. You should simply let the hardware write its depth buffer like normal and then read from it only. If you only read from it you should prevent any performance hits you might take by doing it manually. I have a feeling I'm just misunderstand what you are saying though. Unfortunately I don't know any OpenGL specifics, that's why I haven't responded.

#9 pancakedice   Members   -  Reputation: 159

Like
0Likes
Like

Posted 30 July 2010 - 02:29 AM

Thanks for the help - either I am not getting you or you are not getting my, so I'll try again.

In my first geometry pass I do the following:
I first render to a MRT with a GL_TEXTURE_RECTANGLE as GL_DEPTH_ATTACHMENT and a few buffers as GL_COLOR_ATTACHMENTn. I just let OpenGL store its own depth in the texture, as I do not write to gl_FragDepth. That is, OpenGL stores its own hyperbolic depth just as usual, but in a texture I have attached.

In the second pass I do the following:
Bind my color buffer textures for normals etc., AND bind my depth buffer texture so I can fetch data from it, as I need to reconstruct position.

However, before reconstructing the position I'd like to do a pass (or two) where I only write to the stencil buffer so I can use that to discard pixels, so I do not run my expensive light shader more than needed. For this I need to utilize the depth that is written to the texture to do depth testing against backfaces of the light volume (and perhaps also front faces), but I'm not sure how I can do that, as for the second pass I draw directly to the window's default framebuffer.

Did that make it more clear?

Thanks so far.

#10 allingm   Members   -  Reputation: 452

Like
0Likes
Like

Posted 30 July 2010 - 05:35 AM

I see. It seems like we are on the same page, and it seems like you've solved your depth problem. Now you just have deferred rendering stencil volume problems.

Stencil volumes work something like this:
1) Render front face of stencil volume, and increment stencil buffer.
2) Render back face of stencil volume, and decrement stencil buffer.
3) Render light only where stencil volume is 1, and set stencil buffer to 0.

Special case: when camera is inside the stencil volume. I don't remember what I did here at the moment, but for now you can simply render a full screen quad when you are inside.

I don't know the OpenGL specifics, but it might help you to start a new thread.

#11 allingm   Members   -  Reputation: 452

Like
0Likes
Like

Posted 30 July 2010 - 05:37 AM

I forgot to say a couple things.

1) Turn off color writes for stencil volumes.
2) Turn off depth writes for stencil volumes and lights.
3) Turn on depth reads for stencil volumes.
4) Turn off depth reads for lights. (doesn't really matter here)

#12 pancakedice   Members   -  Reputation: 159

Like
0Likes
Like

Posted 30 July 2010 - 06:41 AM

Yeah, that was what I was thinking. However, "solved" my depth problems is not exactly right, as I still have no clue as to how I am going to utilize the texture as the depth buffer for the second pass, as I am not using MRT for that, due to the fact that I render to the window. Depth testing that I have to use for the stencil generation works with the current depthbuffer unless I am mistaken. In this case as I do not use MRT that would be the window's own depth buffer, and not my custom texture that I just rendered to in order to use the info to reconstruct position.

I might be missing something quite simple, but I'm still not getting anywhere.



#13 venzon   Members   -  Reputation: 256

Like
0Likes
Like

Posted 30 July 2010 - 06:55 AM

Quote:
Original post by dingojohn
For this I need to utilize the depth that is written to the texture to do depth testing against backfaces of the light volume (and perhaps also front faces), but I'm not sure how I can do that, as for the second pass I draw directly to the window's default framebuffer.


To do depth testing using a depth texture that you've previously written to, you need to attach it to your FBO. As far as I know, you can't attach depth textures to the special framebuffer FBO 0. Your options are:
1) draw your second pass to an FBO with your depth texture attached (which will make opengl use it for depth testing) and also a new color texture attached, then do a third full-screen pass to simply write the color texture to the framebuffer. This is what I do, and it's when I apply my tonemapping.
2) insert a full screen pass before your second pass. In this new pass, draw to the special framebuffer FBO 0 and you read from your depth texture and in your fragment shader set the fragment's depth to what you read from your depth texture.

I'd strongly recommend option 1 because then you have you scene's color in a texture, which is useful for all kinds of post-processing.

#14 pancakedice   Members   -  Reputation: 159

Like
0Likes
Like

Posted 30 July 2010 - 07:30 AM

Thanks for the info. There's flying some rate++ towards you!

I'll try this now and hopefully everything will work out brilliantly.

Thanks a bunch!




Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS