[Luigi__]

I found this: http://web.archive.org/web/20070624082603/www.hiend3d.com/smartflt.html while searching about HQ2X algorithm.

This is smart texture filtering. Basically, it takes each block of 4 texels and filters it differently depending on its pattern, in order to keep borders sharp and achieve smoothness inside the borders. There are 14 possible patterns:

Pattern 0: linear/bilinear/whatever smoothing on the 4 pixels
Pattern 1: smoothing on the top pixels, smoothing on the bottom pixels
And so on...

Here are the results with different filtering methods:

Nearest filtering:

Bilinear filtering:

Smart filtering:

I thought it may be interesting to implement a filtering method like that, which can work on any image size, on DeSmuME's window, rather than all those filters which only work with fixed sizes (like HQ2X, HQ3X, HQ4X...).
And maybe later on it can get implemented in the Soft rasterizer, who knows...

Your opinions are welcome.

 

[Engeljaeger]

looks good! though I could imagine that it doesn't recognize all the places it should so that in the end maybe you have some crisp graphics mixed with some blurry ones if you get my meaning...on the other hand I have no knowledge of these things so I'm just waitin' to see it in action ^^

 

[nanaya]

Kewl, I'd like it to be implemented... I can't help on the technical detail though. :heh:

 

[zeromus_]

More to the point, has anyone ever seen this implemented in C instead of a huge heap of ugly asm

 

[Luigi__]

In the worst case, we can just take the interesting part in the asm code

That's the reason why it can't be implemented easily. It will have to be reimplemented. Atleast, we know how it works, which is better than just having the asm and knowing nothing

 

[Squall-Leonhart]

zeromus, shush your mouth D: lol

ASM filters perform far more efficiently then C.
....if you really want C though, i know someone who can pull it off....

though i want this filter in VBA now

 

[Luigi__]

Squall-Leonhart said it all, ASM is faster anyway
And I don't see a problem in just putting the file into desmume, seeing as some other parts (like the input dialog) are taken from elsewhere too...

Edit: I'm wrong. There's a problem. The thing uses extra texel information to locate the borders better.

 

[Kein]

Oh shi~
Looks really great.

 

[Squall-Leonhart]

it requires pre processed images, so it'd only be beneficial on 2d games.

 

[shashClp]

Checking the files, porting it to C would be painful due to mmx usage, thus requiring using intrinsics

 

[Luigi__]

That's why the asm implementation is not good for our usage.

Therefore, if I implement it, I do it from scratch. I'd also make it even smarter by caring about the 4x4 blocks around the block being filtered...

 

[zeromus_]

It does not necessarily require preprocessing. And even if it is implemented with preprocessing, it could easily be done in desmume for textures for 3d due to the texture manager being a very modular system (reused entirely by the ogl and soft rasterizer).

 

[Squall-Leonhart]

>.> i'll smack that gigaherz in the mouth for lieing to me.

 

[zeromus_]

I mean, obviously its more efficient to preprocess, but look at that guy. its 2001 and he coded that algorithm in asm instead of making a reference implementation which is comprehensible. Of course he is going to go out of his way to make it is complicated as possible by preprocessing and making a custom pixel format.

 

[Engeljaeger]

it requires pre processed images, so it'd only be beneficial on 2d games.

...which wouldn't be so bad considering the number of DS games in 2D

 

[zeromus_]

itll probably look like ass, but itll also be trivial to find out once someone has coded it in C

 

[nanaya]

I found some documentation about image interpolation while browsing doom9 forum, something about Edge-Directed Interpolation, some of its :

Edge-Directed Interpolation
http://www.comp.leeds.ac.uk/bmvc2008/proceedings/papers/43.pdf
http://www.eurasip.org/Proceedings/Eusipco/Eusipco2009/contents/papers/1569192778.pdf
http://www.eie.polyu.edu.hk/eInfo4Students/msc_thesis/msc06_03.pdf

Hope this help and sorry if it doesn't or irrelevant... :innocent:

 

[Squall-Leonhart]

ugh, i hate interpolation, SaI makes LCD and non 4:3 images look bad.

 

[ccl0]

this would be amazing if this could be implemented. looks vastly better than bilinear filtering