Hi Jon, Daniel, all,

On 9/10/13, Daniel Sabo wrote:

The conversion happens in app/display/gimpdisplayshell-render.c : gimp_display_shell_render() .
Look for the gegl_buffer_get call that converts to babl_format ("cairo-ARGB32").

On 9/11/13, Jon Nordby wrote:

Quoting Jon's suggestion:
(1)call a "GeglBuffer from the projection (or possibly a copy)" and then So how does one "call a GeglBuffer from the projection"?

I said "operate on", not "call". :)

Sorry!

The projection has a GeglBuffer associated with it, see https://git.gnome.org/browse/gimp/tree/app/display/gimpdisplayshell-render.c#n77 That is what the color conversion needs to use as an input, basically*.

I set up a babl/gegl/gimp prefix for working on this issue and wrote a function that creates a hard-coded ICC profile transform:

static void image_to_monitor () {
cmsHTRANSFORM image_to_monitor_transform = NULL; cmsHPROFILE image_space;
cmsHPROFILE monitor_profile;
image_space =
cmsOpenProfileFromFile("/usr/share/color/icc/V4/AllColors-elleV4-g100.icc","r"); monitor_profile =
cmsOpenProfileFromFile("/usr/share/color/icc/monitor-as-qh.icc","r"); image_to_monitor_transform = cmsCreateTransform (image_space, TYPE_RGBA_FLT, monitor_profile, TYPE_RGBA_8, INTENT_RELATIVE_COLORIMETRIC, cmsFLAGS_NOOPTIMIZE | cmsFLAGS_BLACKPOINTCOMPENSATION ); if (image_to_monitor_transform)
{printf("Image_to_monitor_transform WAS created in display-filter-lcms.c.\n");}
else {printf("Image_to_monitor_transform was NOT created in display-filter-lcms.c.\n");}
}

The purpose of this function is (eventually) to be able to convert the projection from the image profile to the monitor profile without bringing in all the overhead that allows Gimp to detect which profile is which. The overhead can be added later. That's my theory, anyway.

To go along with this hard-coded transform, I made a 32-bit floating point image with an alpha channel. So in my dedicated prefix version of Gimp, that's the only image that gets opened.

My first goal was to get the "image_to_monitor" function to work, not to do anything with it, just to get evidence that the transform was created.

I put the "image_to_monitor" function in "display-filter-lcms.c" and added this line - "image_to_monitor ();" - to the function "cdisplay_lcms_convert_surface". Gimp compiles and runs just fine, printing out lines that the transform was indeed created.

I also added the exact same "image_to_monitor" function to the file "gimpdisplayshell-render.c", except modified so that the printf lines say "WAS/was NOT created in gimpdisplayshell-render.c". Again, Gimp compiles and runs with no problems, but no lines are printed from this function because at this point it isn't called by the function "gimp_display_shell_render".

Unfortunately, as soon as I add the line "image_to_monitor ();" to the function "gimp_display_shell_render", then Gimp won't compile, but rather exits with these lines:

make[4]: Leaving directory `/home/elle/code/gimp292/build/gimp/app/gui' Making all in .
make[4]: Entering directory `/home/elle/code/gimp292/build/gimp/app' CCLD gimp-2.9
display/libappdisplay.a(gimpdisplayshell-render.o): In function `image_to_monitor':
/home/elle/code/gimp292/build/gimp/app/display/gimpdisplayshell-render.c:79: undefined reference to `cmsOpenProfileFromFile' /home/elle/code/gimp292/build/gimp/app/display/gimpdisplayshell-render.c:80: undefined reference to `cmsOpenProfileFromFile' /home/elle/code/gimp292/build/gimp/app/display/gimpdisplayshell-render.c:81: undefined reference to `cmsCreateTransform'

Any ideas as to why Gimp won't compile when I add the line "image_to_monitor ();" to the function "gimp_display_shell_render"? It's pretty obvious that it's not finding lcms2.h. But I did add the following "#include" lines to the top of gimpdisplayshell-render.c:

#include /* lcms.h uses the "inline" keyword */ #include
#include

Suggestions regarding other/better ways to approach the problem are also welcome!

Elle

Hmm, I managed to get past the problem of not linking to lcms2. The solution was to compile Gimp using MAKEOPTS="-j3" CFLAGS=-Wl,--no-as-needed LDFLAGS=-llcms2 ./autogen.sh . . . not that I know why that seems to have worked.

On to the next step.

http://ninedegreesbelow.com - articles on color management & open
source photography

I've made progress getting a linear gamma image to display without posterization in the shadows. I put up a temporary web page with the modified code and a picture:
http://ninedegreesbelow.com/temp/convert-buffer-before-cairo.html

The problem at this point is that the image won't display properly until something like doing a very small levels correction forces a screen redraw. After forcing a screen redraw, the image is displayed without any posterization, but with magenta lines (outlining the tiles?). The screen redraw lasts until the level dialog is closed.

I think the problem is that I'm not properly merging and updating "buffer" after the hard-coded transform. The corresponding code from the lcms.c file uses layer buffers, which seems not applicable to a projection:
gimp_drawable_merge_shadow (layer_id, TRUE); gimp_drawable_update (layer_id, 0, 0, layer_width, layer_height);

If I can get the screen to consistently display properly, then the next step would be to merge the code in "modules/display-filter-lcms.c" with the code in "app/display/gimpdisplayshell-render.c". But I'm somewhat at a stand-still until serendipity or someone with more knowledge about gegl buffers can solve the problem of getting the screen to display all the way all the time correctly rather than only after using levels to force a redraw.

I've been searching the Gimp code base looking for that "serendipitous" example code to follow, and came up with "gimp_projection_flush_now (projection);" which seemed to help, but only a little.

Elle

http://ninedegreesbelow.com

On 9/12/13, Jon Nordby wrote:

On 12 September 2013 16:08, Elle Stone wrote:

The problem at this point is that the image won't display properly until something like doing a very small levels correction forces a screen redraw. After forcing a screen redraw, the image is displayed without any posterization, but with magenta lines (outlining the tiles?). The screen redraw lasts until the level dialog is closed.

I think the problem is that I'm not properly merging and updating "buffer" after the hard-coded transform. The corresponding code from the lcms.c file uses layer buffers, which seems not applicable to a projection:
gimp_drawable_merge_shadow (layer_id, TRUE); gimp_drawable_update (layer_id, 0, 0, layer_width, layer_height);

I do not know the GimpDisplayShell code well, but try to just read out data from the projection GeglBuffer instead of modifying it.

I created a copy of "buffer" and converted it to the monitor profile, with the same results. I'm pretty sure the problem is what happens to the buffer after it's been used. It doesn't quietly disappear. In a small test image the code that does the transform gets executed 11 times per anything that changes the screen. It takes up "pipes" and a larger image eventually crashes Gimp ("unable to open pipe: Too many open files").

And
instead of the the "gegl_buffer_get (buffer, ... "cairo-ARGB32", ... data ...)" that you have marked, do the lcms transform such that the 8bit ready-for-display ends up in the "data" buffer.

That code is the pre-existing code that sends the buffer to cairo, so I haven't tried to modify it. There's probably a way to get "buffer" to be converted from the image color space at 32f to the monitor profile at 8i in one fell swoop - I'm pretty sure lcms can do the conversion in one fell swoop but I don't have the gegl buffers set up correctly. But "gegl_buffer_get (buffer, ... "cairo-ARGB32", ..." does convert "buffer" to 8-bits.

Also, can you please post your changes as a (git formatted) diff? It is much easier to read and apply for another contributor trying to help you out.

I posted the git patch, the two ICC profiles in the hard-coded transform, and a test image to
http://ninedegreesbelow.com/temp/convert-buffer-before-cairo.html. I hope I did the git patch correctly! The git patch (but not the profiles or test image) is also attached to this email.

Jon Nordby - www.jonnor.com

Thanks! Jon, for taking an interest in this project.

Elle

http://ninedegreesbelow.com

Hi Daniel,

Thanks! for the input. I'm pondering your suggestions about the code and I'll check in with you and Mitch on IRC in the next couple of days.

http://ninedegreesbelow.com
Just because it's a standard, doesn't mean it's right.

Massimo made a patch that was so much better than my poor effort. His patch worked really well. Then Mitch did a whole bunch of code writing and now Gimp from git can properly display linear gamma images.

Thank you Massimo! Thank you Mitch!