5
$\begingroup$

I'm in the process of working out how to pack all the information I need for a Physically Based Deferred Renderer into a G-Buffer without using an obscene amount of render targets.

What I have so far is 4 3-part vectors:

  • Albedo/Diffuse
  • Normal
  • Tangent
  • Position

And 4 single components

  • Metallic
  • Roughness
  • Height
  • Ambient Occlusion

A naive approach is to bundle one of the single components into the alpha (fourth) channel with one of the 3-part vectors, which is my current line of investigation. However, given that four 4-channel full precision floating point render targets isn't small I understand it's common to use half precision and even smaller representations to be more memory conscious.

What I'm asking is: which components can I safely cut precision down on without losing quality, and by how much?

$\endgroup$
4
$\begingroup$

First of all, you don't need position in the G-buffer at all. The position of a pixel can be reconstructed from the depth buffer, knowing the camera setup and the pixel's screen-space xy position. So you can get rid of that whole buffer.

Also, you don't ordinarily need tangent vectors in the G-buffer either. They're only needed for converting normal maps from tangent space, and for parallax mapping; these would be done during the G-buffer fill pass (when you have tangents from the mesh you're rendering), and the G-buffer would only store normals in world or view space.

Material properties like colors, roughness, and metallic are usually just 8-bit values in the G-buffer, since they're sourced from 8-bit textures. Same for AO.

Height is also not needed in the G-buffer unless you're going to be doing some kind of multi-pass blending that depends on it, but if you do need it, 8 bits is probably enough for that too.

Normals can be benefit from being stored as 16-bit values rather than 8-bit. Half-float is okay, but 16-bit fixed-point is even better, as it gives you more uniform precision across all orientations (half-float is more precise near the axes and loses some precision away from them). Moreover, you can cut them from 3 components down to 2 using octahedral mapping.

So, at the end of the day, a minimal G-buffer might look like:

  • Material color + metallic: RGBA8
  • Octahedral world-space normal + roughness + AO: RGBA16

and that's all! Only 12 bytes per pixel.

Alternatively, you could use an RG16 buffer for the normals, and move roughness + AO into a separate 8-bit buffer. That would give you some room to grow should you eventually need more G-buffer components of either 8-bit or 16-bit sizes.

| improve this answer | |
$\endgroup$
  • 1
    $\begingroup$ Also, metalness is often a binary value, and assuming that you will not use roughness gradients (so as to see the impact of reduced precision), you could store metalness and roughness in a single 8 bit channel (1 bit for metalness and 7 for roughness). $\endgroup$ – Kostas Anagnostou Apr 7 '17 at 15:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.