Christoph Hormann schrieb:

>> - Image-format input files used immediately in certain language 
>> constructs obviously indicating that they are mere data containers 
>> (e.g. height fields) shall be exempt from the aforementioned 
>> conversion, and the raw encoded values used instead. (Note however 
>> that this does not apply to indirect use of input image files.)
> 
> It would OTOH be nice if POV-Ray would treat images the same way 
> everywhere (it does not in 3.6 either but that's a different story).

Well, I think there's some reason to both sides; with the per-file gamma 
setting it becomes less of a problem at any rate.


>> - Non-paletted integer images are kept in "native" format without 
>> applying gamma adjustment, in order to save memory. Gamma correction 
>> is instead applied at render time, using a pre-computed lookup table 
>> in hope of gaining a bit of a speed advantage. (Lookup tables are 
>> shared between images with the same gamma curve of course.)
> 
> A dynamic caching scheme for converted values would be an alternative - 
> significantly more complex to implement of course.

... and quite definitely more overhead than gain for an 8-bit/channel 
image, and possibly so as well for a 16-bit/channel image (except for 
images so small that it would become worthwhile to convert them to float 
right away at load time). It might also be questionable whether it would 
be faster than just recomputing the value.

> As long as this is only about gamma correction this is probably 
> unnecessary but with color space conversions on-the-fly conversion might 
> be too expensive.

For some "norm" color spaces, I think the per-color lookup table 
approach might still be both possible and worthwhile - especially for 
8-bit/channel images - if separate lookup tables are used for each color 
channel, and each entry would store an RGB color instead of a scalar; 
those would then be just added - sort of a shortcat halfway through the 
matrix transformation found in various color space transformations.

Of course for full-fledged device color profiles this wouldn't be 
sufficient.

Post a reply to this message

clipka wrote:

> - POV-Ray shall always operate in a linear RGB color space based on 
> light intensity (as opposed to perceived brightness), to properly 
> compute physical effects. (In the long run, this should also include a 
> clear definition of the RGB primaries and white point, but at present 
> this definition is limited to gamma issues.)
> 
> - In General, all image-format input files sall be converted to 
> POV-Ray's operating color space before any further processing. (In the 
> long run, this should include full color profile handling; however, at 
> present it is limited to gamma correction.)
> 

So meanwhile you do agree with me that POV-Ray should have "in the long 
run" a well defined RGB color space while I do remember a few months ago 
we had some struggle about this issue ;)


> - For file formats with an official specification...
 > [SNIP (a lot)]

Sounds all very good to me.


> * Implementation
 > [SNIP]
> - Non-paletted integer images are kept in "native" format without 
> applying gamma adjustment, in order to save memory. Gamma correction is 
> instead applied at render time, using a pre-computed lookup table in 
> hope of gaining a bit of a speed advantage. (Lookup tables are shared 
> between images with the same gamma curve of course.)
> 

I already wanted to ask you about this - as at some point it seemed to 
me you intend to store all data in floating point format - but then I 
did completely forget about it.
While I think it is a very good idea to use lut's for 8 bps I do believe 
it should not be done for 16 bps as they are meant to apply things like 
inverse gamma correction without *any* visual impact. But this is not an 
important issue.


And BTW I've rewritten tiff.cpp from scratch and the most simple and 
basic cases is already working: RGB data that is ordered interleaved 
(and not in separate color planes) arranged in strips (and not tiles) 
and uses unsigned 8 or 16 bit integers (not signed ones).
The nice thing is that full floating point HDR TIFF support for 16bit 
(the same 'half' format as used by OpenEXR), 24bit (yes, such a thing 
exists) and 32bit (the 'normal' float) is already working.
And I intend to implement TIFF *output* for 32bit floats as I think it 
would be nice to have this feature available especially within an 
already well defined image file format.
So sooner or later I need to know more details about your implementation 
of the file gamma correction but most probably I will not finish the new 
TIFF implementation befor the next beta is out...


-Ive

Post a reply to this message

Ive schrieb:
> So meanwhile you do agree with me that POV-Ray should have "in the long 
> run" a well defined RGB color space while I do remember a few months ago 
> we had some struggle about this issue ;)

Ah yes? Did we?

If you're thinking of statements from me like "what's the point in 
having perfect ICC profile handling if POV-Ray gets the gamma wrong" 
(about a workaround I had described for the gamma issues with POV-Ray, 
involving an external tool that according to you doesn't do proper ICC 
handling), and consider those an argument about ICC profiles, then yes, 
in that case I guess we did :-)


> I already wanted to ask you about this - as at some point it seemed to 
> me you intend to store all data in floating point format - but then I 
> did completely forget about it.
> While I think it is a very good idea to use lut's for 8 bps I do believe 
> it should not be done for 16 bps as they are meant to apply things like 
> inverse gamma correction without *any* visual impact. But this is not an 
> important issue.

Indeed it isn't: The framework has been set up now, and adjustments to 
the handling should be quite easy. Including stuff like, checking for 
the file size and depending on that choosing whether to convert directly 
to float or use lookup.

> And BTW I've rewritten tiff.cpp from scratch and the most simple and 
> basic cases is already working: RGB data that is ordered interleaved 
> (and not in separate color planes) arranged in strips (and not tiles) 
> and uses unsigned 8 or 16 bit integers (not signed ones).
> The nice thing is that full floating point HDR TIFF support for 16bit 
> (the same 'half' format as used by OpenEXR), 24bit (yes, such a thing 
> exists) and 32bit (the 'normal' float) is already working.
> And I intend to implement TIFF *output* for 32bit floats as I think it 
> would be nice to have this feature available especially within an 
> already well defined image file format.

Great to hear those things!

> So sooner or later I need to know more details about your implementation 
> of the file gamma correction but most probably I will not finish the new 
> TIFF implementation befor the next beta is out...

It's actually quite easy as far as the image reading code is concerned. 
Nothing that would require a dramatic change in image reading code.

Post a reply to this message

clipka schrieb:
> 
> ... and quite definitely more overhead than gain for an 8-bit/channel 
> image, and possibly so as well for a 16-bit/channel image (except for 
> images so small that it would become worthwhile to convert them to float 
> right away at load time). It might also be questionable whether it would 
> be faster than just recomputing the value.

Could be.

One field where i consider performance to be possibly critical is use of 
images in isosurfaces.  This leads to a real lot of image evaluations. 
There however you nearly always use a linear gamma curve.  So a special 
case for images where file_gamma is exactly 1.0 would be a good idea.

-- Christoph

Post a reply to this message

Christoph Hormann schrieb:

> One field where i consider performance to be possibly critical is use of 
> images in isosurfaces.  This leads to a real lot of image evaluations. 
> There however you nearly always use a linear gamma curve.  So a special 
> case for images where file_gamma is exactly 1.0 would be a good idea.

That's already implemented :-)

In case the effective decoding gamma curve happens to be neutral, the 
file readers choose an image container that does not do gamma lookup at 
all, and just does the standard "/255.0f" decoding instead.

Post a reply to this message

Ive <"ive### [at] lilysoftorg"> wrote:
> While I think it is a very good idea to use lut's for 8 bps I do believe
> it should not be done for 16 bps as they are meant to apply things like
> inverse gamma correction without *any* visual impact. But this is not an
> important issue.

What if it's a lookup table that maps a 16-bit integer into a 32-bit float?
There wouldn't be any visual impact there...

Post a reply to this message

Nicolas Alvarez schrieb:
> Ive <"ive### [at] lilysoftorg"> wrote:
>> While I think it is a very good idea to use lut's for 8 bps I do believe
>> it should not be done for 16 bps as they are meant to apply things like
>> inverse gamma correction without *any* visual impact. But this is not an
>> important issue.
> 
> What if it's a lookup table that maps a 16-bit integer into a 32-bit float?
> There wouldn't be any visual impact there...

I think what Ive means is that 16 bit per color channel should already 
provide enough "headroom" to do the gamma correction in that format.

Post a reply to this message