Re: [Algorithms] Texture mip streaming prioritization

[Sebastian S. <seb...@gm...>]

On Mon, Jul 22, 2013 at 12:16 AM, Josh Green <in...@gm...> wrote:

> Hi There,
>
> I'm currently working on a texture streaming system and looking into ways
> to prioritize which mips on which textures should be loaded.
> I was hoping that people on this list may have experience with these
> algorithms that they could share?
>
> Currently I know which mips need to be loaded in order to render a scene
> without degradation in image quality.
> I would now like to apply a memory budget to textures, and make decisions
> about which mips of which textures should be loaded over others.
>
> My current line of thinking involves comparing cost / benefit of loading
> each individual mip.
> i.e.
>
> How many pixels would be affected by a particular mip level being loaded
> versus how much memory would that mip level use?
>
> This becomes Priority = Benefit / Cost
>
> I'd then sort for highest priority and assign memory budget to those mips
> at the top of the list.
>
> Any thoughts? comments? Alternatives that work better? Alternatives that
> have worked enough?
>

I would bake in some kind of factor that relates to "quality improvement"
as well. If you have a 4k eye ball texture and it's currently at MIP 2
(1Kx1K), it probably won't gain much from loading another MIP even if it's
covering a lot of pixels. So you need to compute how far off the pixel
currently is from the *ideal* MIP level, and improve pixels that are far
off first (weighted by pixel count).


Assuming you only load one MIP level at a time (i.e. if MIP N is currently
loaded, you'll only consider N-1), maybe something like:

(desired_mip - current_mip) * num_pixels / memory_usage, add tuning knobs
as needed...

That first factor relates to how far "off" a pixel currently is from its
ideal MIP level, i.e. how much in need of improvement those pixels are.

-- 
Sebastian Sylvan