Comments on Icare3D Blog: Fast and Accurate Single-Pass A-Buffer using OpenG...

I compiled it on GCC4/Linux. Some notes: - in Matr...

2010-06-27T08:43:52.964+02:00

I compiled it on GCC4/Linux. Some notes:
- in Matrix4.h, "static Mat4 reflection(const Vector4 &plane)" does not compile because "v" is undeclared
- I had to include and for fprintf, stderr and strcmp to be declared in ShadersManagment.cpp
- it should be "Management" and not "Managment" ;)

Great technique BTW, too bad I only have a GeForce 8600M GT :/
When running glewinfo, I get "OK [MISSING]" for the extensions you use, which means the functions are here but not in GL_EXTENSIONS.
Do you know if these extensions could work anyway on my hardware, or if they are just present because the driver is the same for Fermi and G80 hardware?

I think you need an R256 drivers to get OpenGL 4.0...

2010-06-15T15:50:03.067+02:00

I think you need an R256 drivers to get OpenGL 4.0 support. You should try the new 257.21

freeglut (ABufferGL4.exe): Unable to create OpenGL...

2010-06-15T15:37:26.093+02:00

freeglut (ABufferGL4.exe): Unable to create OpenGL 4.0 context (flags 2, profile 1)

drivers: 197.75

Hi Ben, a 470 should run it, do you get an error m...

2010-06-14T22:22:17.962+02:00

Hi Ben, a 470 should run it, do you get an error message on the console ? What driver are you using ?

Is there a list of cards this will run on? My 470 ...

2010-06-14T20:29:17.320+02:00

Is there a list of cards this will run on? My 470 won't run the demo.

@Rex Guo : It does make sense. It is an interestin...

2010-06-11T11:15:16.826+02:00

@Rex Guo : It does make sense. It is an interesting application of all A-Buffer approaches I think.

@pixelstoomany : Thanks for the link !

@id: Thanks :-)

@Todd: Thanks also for the link, I will read this !

The AMD method uses a per pixel linked list and is...

2010-06-10T21:43:27.691+02:00

The AMD method uses a per pixel linked list and is described in a GDC 2010 presentation. I assume the EGSR presentation will have more examples/uses.

http://developer.amd.com/gpu_assets/OIT%20and%20Indirect%20Illumination%20using%20DX11%20Linked%20Lists_forweb.ppsx

Awesome work Cyril. Always nice to see someone us...

2010-06-10T21:18:41.227+02:00

Awesome work Cyril. Always nice to see someone using OpenGL 4.0 with the core profile. :) Sure, there are plenty of other OIT algorithms out there, but A-Buffer is nice and simple. I love it!

AMD method uses per pixel linked lists (and it'...

2010-06-10T19:02:19.585+02:00

AMD method uses per pixel linked lists (and it's very fast), they have a paper at EGSR that explains their method and a few applications.

Translucent shadow maps are possible (I have implemented a few different solutions) but generally very slow to sample. There are different and imho better ways to approach that problem and we have a paper at EGSR to show an alternative approach.

See the "Shadow and Order Independent Transparency" session:
http://kesen.realtimerendering.com/egsr2010Papers.htm

That's very nice work - keep it up! It looks ...

2010-06-10T16:44:44.385+02:00

That's very nice work - keep it up!

It looks like this is something that can be extended
to implement translucent shadow maps? During the depth
pass, things are rendered exactly like what you've
described here (except that the final pass for merging
is skipped). During the render pass, a Z-search is
done on the A-buffer to look for where on the list
the view space pixel is. If there're samples between
this pixel and the light (i.e. shadowed by one or more
layers), then all the samples are alpha-blended to give
the final colour.

Not sure if the above makes any sense.. :-)

I dis not know that technique, do you have a refer...

2010-06-10T12:42:25.769+02:00

I dis not know that technique, do you have a reference to a paper or a presentation ?
Thanks !

Erm... Actually the AMD method doesn't have a ...

2010-06-10T12:37:36.178+02:00

Erm... Actually the AMD method doesn't have a fixed number of fragment per pixel.
That's where their solution is slighly more interesting. A single pixel can contain 100 fragments even if the average fragment per pixel is 16 for example. It meat that we expect that some pixel will require less than 16 samples which is likely to happen. I also suspect that the AMD method is faster but this is difficult to actually check.

Well this is still a good job but I am looking forward for the further result of your research! :p

Yes there do is a fixed maximum amount of fragment...

2010-06-09T22:06:34.861+02:00

Yes there do is a fixed maximum amount of fragments per pixel, but the actual number is maintained dynamically (that's the per-pixel counter updated each time a fragment is written).

Does your technique store a fixed amount of fragme...

2010-06-09T21:59:42.557+02:00

Does your technique store a fixed amount of fragments per pixel? Or is it flexible and it allows you to store a variable number of fragments per pixel? (similarly to what AMD does in their OIT demo..)

Hi Christophe ! You are right you have to allocate...

2010-06-09T20:25:59.378+02:00

Hi Christophe !
You are right you have to allocate a big buffer at the beginning with the maximum number of layers on every pixels. That the main problem with this technique, but it was also the case with previous techniques (and as you said, here it does not consume any bandwidth).
Can you elaborate a little bit more on old days solutions you are talking about ?

Niceee!! This is actually probably more an OpenGL...

2010-06-09T19:47:23.616+02:00

Niceee!!

This is actually probably more an OpenGL 4.1 method (or beyond)!
It doesn't work with AMD drivers like you expected.

I might I miss something but you have to allocate a big buffer at the beginning for the written fragments? A bit like the AMD order independent transparency but with a 2D texture array instead of RWBuffer?

One thing which feels not so practical is the amount of memory required for the A-Buffer even in areas where actually no fragment are written. Yes, at least it doesn't consume any bandwidth...

There are solutions explored back in the old days where no extra memory and no extra bandwidth were required but that never made it in nVidia and AMD chips... I don't really understand.