Not many can claim 25 years on the Internet! Join us in celebrating this milestone. Learn more about our history, and thank you for being a part of our community!
Hi!
I have ported Light Indexed Deferred Lighting from Original Demo: to Direct3D12 Version and OpenGL core profile version
I have bug during final shading pass for Direct3D12 version:
And OpenGL Version with some modification:
2) Simple light source animation
3) Also I tested an optional Direct3D clip control mode
I Guess may be Problem with Depth Bound Test. I need fixing for Direct3D12 only. I can provide additional information
Hov to fix it ?
The problem seems to happen only where many lights affect the surface? So maybe just some buffer overrun?
Hi @joej thank you for your reply.
JoeJ said:
The problem seems to happen only where many lights affect the surface? So maybe just some buffer overrun?
Yes, problems happens during affecting multiple lights to pixel
But I dont't know about buffer overrun. Lights source geometry is drawing to RGBA8 RenderTarget. What Burrer do you mean?
AndreyVK_D3D said:
What Burrer do you mean?
I just thought you would do something like making a list of lights per tile, and this could break if the list becomes too large.
I forgot how indexed deferred is meant to work. You could explain, but you would only get some other guesses in return. ; )
JoeJ said:
I just thought you would do something like making a list of lights per tile, and this could break if the list becomes too large.
Light Indexed Deferred Lighting(LIDR) doesn't use Light list per tile/cluster like as Forward+/Cluster Shading
JoeJ said:
I forgot how indexed deferred is meant to work. You could explain, but you would only get some other guesses in return. ; )
1) DepthTest ON/Depth Write OFF. This stage is using constant Blending (D3D12_BLEND_ONE, D3D12_BLEND_BLEND_FACTOR) with 0.25, 0.25 0.25 0.25 params https://docs.microsoft.com/en-us/windows/win32/api/d3d12/nf-d3d12-id3d12graphicscommandlisomsetblendfactor So, We can mix only 4 Lights per pixel.
2) Light sources should be clipped using Stencil or Depth Bounds Test(I Use DBT)
3. Final shading pass: Extract 4 light Indices from RGBA Read Each Light source from ConstantBuffer and calculate final color.
AndreyVK_D3D said:
So, We can mix only 4 Lights per pixel.
And so you get artifacts if more lights affect it, and there is nothing you can do to fix it, i guess?
Did read a bit in this paper, which seems to confirm this limitation: https://github.com/dtrebilco/lightindexed-deferredrender/blob/master/LightIndexedDeferredLighting1.1.pdf
In the paper they propose to do a second pass if such cases happen, mentioned in the section ‘Light Index Packing - CPU sorting’. But that's heavy cost, and the same problem comes back if light count exceeds 8.
Maybe some stochastic solution could be worked out as well, so the artifacts become temporal noise and at least acceptable. Though, some Forward+ would be both easier and more robust i think.
I also saw @mjp used the technique to compare performance. Maybe he can comment when he comes by…
https://therealmjp.github.io/posts/light-indexed-deferred-rendering/
JoeJ said:
In the paper they propose to do a second pass if such cases happen, mentioned in the section ‘Light Index Packing - CPU sorting’. But that's heavy cost, and the same problem comes back if light count exceeds 8.
Ok, This is an important information.
JoeJ said:
Maybe some stochastic solution could be worked out as well, so the artifacts become temporal noise and at least acceptable. Though, some Forward+ would be both easier and more robust i think.
Yes, I want to swith to more modern Technique like as Forward+/cluster shading. But I want to know about this artefacts.
JoeJ said:
I also saw @mjp used the technique to compare performance. Maybe he can comment when he comes by…
Yes, I read this article long times ago during researching LIDR.
Thank You.
FYI what I implemented in that old blog post was what (later) became known as Forward+. At the time to me it seemed like a variant on light indexed deferred rendering since you were building lists of lights in screen space (just per-tile instead of per-pixel) and so I called it by that name. But I think I was the only one to make that connection, and soon after the Forward+ name came around and took over.
