YCoCg DXT5 support

[castano]

YCoCg DXT5 support

[GoogleCodeExporter]

Original comment by [email protected] on 4 Feb 2008 at 7:59

• Added labels: Milestone-Release2.1

[GoogleCodeExporter]

In order to support that, the following things need to be done:

- Add ColorTransform_YCoCg.
- Add support for BC3 compression with AlphaMode_None.

The user would then specify the following options:

inputOptions.setAlphaMode(AlphaMode_None);
inputOptions.setColorTransform(ColorTransform_YCoCg);

compressionOptions.setFormat(Format_DXT5);
compressionOptions.setColorWeights(1, 1, 0, 1);

Just like with BC1n this mode could also be optimized so that the DXT1 
compressor
only fits the two first color components.

Alternatively we could simply add a new format:

compressionOptions.setFormat(Format_DXT5_YCoCg);

Original comment by [email protected] on 1 May 2008 at 9:48

[GoogleCodeExporter]

We could also name the color transform 'ColorTransform_CoCgSY' to make clear 
that the
blue component encodes the scale, and that the alpha encodes the luminance.

Original comment by [email protected] on 1 May 2008 at 10:03

[GoogleCodeExporter]

I've checked in an implementation of ColorTransform_YCoCg provided by Jim 
Tilander.

Original comment by [email protected] on 6 May 2008 at 11:20

• Changed state: Started

[GoogleCodeExporter]

Original comment by [email protected] on 11 Oct 2008 at 7:29

• Added labels: Priority-High
• Removed labels: Priority-Medium

[GoogleCodeExporter]

I use the algorithm invented at 
http://developer.nvidia.com/object/real-time-ycocg-dxt-
compression.html,in many cases it works perfect,but in some case it also have 
color 
bleeding artifacts.could anybody give me some advice? 

Original comment by [email protected] on 16 Jan 2010 at 8:50

Attachments:

• uncompressed.bmp
• compressed.bmp

[GoogleCodeExporter]

duge0413, this is not the place for general questions about our paper. A better 
idea is to mail us directly, or use the developer mailing list. In any case, to 
answer your question:

DXT is in general a lossy compression algorithm, that is, artifacts are to be 
expected, and generally present themselves as block compression artifacts, this 
is most noticeable if you use a real-time compression such as the one proposed 
in our paper. YCoCg-DXT5 reduces these artifacts significantly compared to 
DXT1, but does not remove them entirely. Color bleeding usually occur when the 
chroma cannot be approximated accurately with a line, this is probably the case 
in the subpixel antialiased fonts that your screenshot shows.

Original comment by [email protected] on 14 Jun 2010 at 11:27

[GoogleCodeExporter]

I see this is checked in about 2 years ago... is it still planned to be 
released with version 2.1?

Original comment by [email protected] on 23 Jul 2010 at 5:31

[GoogleCodeExporter]

That's still the plan. Sorry, it was a mistake to make so many breaking changes 
in 2.1, that has delayed things way too long, and devote too little time to 
NVTT to make progress at a faster rate. 2.1 should hopefully be released 
sometimes this summer.

Original comment by [email protected] on 26 Jul 2010 at 8:19

[GoogleCodeExporter]

Hi, Ignacio.

How close are you to release Release 2.1? I'm about to implement this feature 
in our texture pipeline which's already using NVTT 2.0.6.  So if this is gonna 
be released anytime soon, I'd just wait for it.

Original comment by [email protected] on 29 Oct 2010 at 1:03

[GoogleCodeExporter]

I've actually been making some good progress toward the 2.1 release recently. 
In the end I'm changing a lot more things that originally planned, which is why 
this has taken so long.

I think I'll have an alpha release pretty soon, probably sometime next week.

Original comment by [email protected] on 29 Oct 2010 at 1:43

[GoogleCodeExporter]

oh that's very cool.. it's a good thing that I checked with you.  My lead, Ian, 
will be very happy, too. (You worked with him before) :D

nvtt tool makes my job really easy, and I really appreciate what you are doing 
for this tool even after you left nvidia :)

Original comment by [email protected] on 29 Oct 2010 at 2:08

[GoogleCodeExporter]

Hmm... I've worked with a few Ian's, but I suppose you are at Relic and are 
talking about Ian Thomson, right?

We use NVTT ourselves in The Witness, so I actually have more time and 
motivation to keep it in good shape now than I used to have at nvidia. We have 
been using 2.1 internally for a while, so it's usable, but only because I know 
what's implemented and what's missing, and I can fix stuff easily as I need it.

Original comment by [email protected] on 1 Nov 2010 at 7:38

[GoogleCodeExporter]

That's right.. Ian T. I'm glad you actually know proper spelling of his last 
name.. :-) (people often put 'p between m and s.)  It's also good that 2.1 is 
already somewhat tested code.

Actually one question about this YCoCg transformation.  Does it convert sRGB 
images to linear before encoding as YCoCg?

Original comment by [email protected] on 1 Nov 2010 at 8:21

[GoogleCodeExporter]

That was a common problem with all the color transforms, it wasn't clear 
whether they should be done in linear or sRGB space, before or after 
mipmapping, and the order of multiple color transforms wasn't obvious either. 

Several people suggested adding an imperative API and that has become the major 
feature in 2.1.  So, instead of providing all the image data and manipulation 
options upfront through the InputOptions API in a declarative way, there's a 
new TexImage class that represents an array of images or faces that you can 
manipulate directly and allows you to implement things in an imperative way.

The resulting code is a bit more verbose, but much more concise and internally 
the code is cleaner and doesn't have to make assumptions about what you want. 
Also, the old InputOptions API is implemented on top of TexImage and the 
resulting code is a bit more simple and easier to maintain.

Original comment by [email protected] on 1 Nov 2010 at 11:45

[GoogleCodeExporter]

Oh, I may write it right, but as he can attest, I always pronounced it wrong. 
:-)

Original comment by [email protected] on 1 Nov 2010 at 11:46

[GoogleCodeExporter]

And to answer your last question, YCoCg is a linear transform, so you can 
theoretically do the sRGB transform before or after. If you are using YCoCg you 
cannot use the SRGB sampler state/texture type, because Y is stored in the 
alpha and that's not automatically transformed, but using sRGB space is still a 
good idea. A simple cheap approximation is to only square the color component. 
Or you could use a combination: sRGB for the chrominance using the texture 
hardware, and pow(2) for the luminance in the shader.

Exposing all these variations in the declarative API would be a nightmare, but 
with the imperative API is trivial.

Original comment by [email protected] on 1 Nov 2010 at 11:55

[GoogleCodeExporter]

Yeah that was why I asked.  I kinda wanted to have YCoCg in linear space(if it 
doesn't lose the quality with dark pixels) so that I can avoid doing manual 
gamma correction in shader.  But now I think linear YCoCg would suffer from 
quality loss, so I guess I have no choice but doing manual gamma correction.  

Doing pow(gamma) for the luminance only in shader sounds like a good approach, 
since power is a scalar op on console I believe.

Thanks.  I'm really looking forward to version 2.1 now :)

Original comment by [email protected] on 2 Nov 2010 at 12:06

[GoogleCodeExporter]

OK, I've recently checked in the changes I was talking about. There are still a 
few minor issues, but I think that it should mostly work.

For the 2.1 beta release I'd still like to polish things a bit more and do a 
lot more testing. Note that the YCoCg stuff is completely untested.

I'll be integrating this version into our code base tomorrow, chances are I'll 
catch a few bugs in the process. I'll probably launch the beta sometime after 
that, my wife is having a baby at any time now, so I'll soon be taking a few 
days off.

Original comment by [email protected] on 5 Nov 2010 at 7:19

[GoogleCodeExporter]

that is an excellent & amazing news(having baby part).  Big Gratz.  I was under 
impression that you had a kid already (i might have read it from your blog long 
time ago... or am I misled ).  If s/he's your 2nd child, make sure you still 
pay enough attention to your first child( more than to nvidia texture tools :) 
) so that he doesn't feel like he's left behind!

Hope everything is well with your beautiful and almost-new-born baby! 

Original comment by [email protected] on 5 Nov 2010 at 2:32

[GoogleCodeExporter]

YCoCg has been supported for a while now. One could write a much better 
compressor for YCoCg images, but for the time being, the current implementation 
is good enough.

Original comment by [email protected] on 27 Sep 2011 at 6:45

• Changed state: Fixed