Examples using imgui not displaying correctly on Linux (Mesa, Intel Ivybridge)

[tivek]

On the Linux OpenGL (Mesa) renderer the examples 9, 11, 12, 13, and 14 show only those parts of their scenes which overlap with the imgui area. The rest of the window is black. Debug text is being displayed (blitted?) every frame, but not cleared between swaps. Attached screenshots show the problem clearly. Tested on Mesa Intel-dri (Ivybridge) 9.2.2 as well as the current git master, both on Linux kernel 3.11.6. Right now I am unable to verify on other operating systems or hardware, but will try to test a Windows OGL build soon.

Only the imgui examples seem to be affected, others work correctly. Throwing out imgui makes the scenes render as expected. Together with the debug text being blitted but not cleared, this somehow seems as a scissoring or blending mode problem. However, in ex11 I find that Context::rendererSubmit() does call glDisable(GL_SCISSOR_TEST) for the scene geometry every frame.

I am still only learning the ropes but would like to get to the bottom of this. Any hints what to try next? Of course verification on other OpenGL configurations would clarify if this is somehow a mesa bug.

Thanks,
Tomislav

[bkaradzic]

I have only Sandybridge on Linux (also there are no issues with AMD nor NV GPUs), and everything there works. Can you check your Mesa version?

$ glxinfo | grep OpenGL
OpenGL vendor string: Intel Open Source Technology Center
OpenGL renderer string: Mesa DRI Intel(R) Sandybridge Mobile
OpenGL version string: 3.0 Mesa 9.1.7
OpenGL shading language version string: 1.30

What do you get on your machine?

[tivek]

Thanks for the quick reply. I am now on Mesa 10.0.0_devel.59506-1 (yesterday's git master) but 9.2.2 behaves the same.

$ glxinfo | grep OpenGL
OpenGL vendor string: Intel Open Source Technology Center
OpenGL renderer string: Mesa DRI Intel(R) Ivybridge Mobile 
OpenGL version string: 3.0 Mesa 10.0.0-devel (git-ff337bc)
OpenGL shading language version string: 1.30

Desktop compositing does not make a difference, nor does switching between SNA and UXA. Maybe I should try out Mesa 9.1.7.. Anyway, I'm glad it's not a direct bug in bgfx.

[bkaradzic]

I have confirmation that this is happening on IvyBridge + Linux.

[tivek]

Thank you. I will try to get a minimal program which reproduces the bug.

[bkaradzic]

@tivek I tested on Haswell as well, and I don't see problem. The difference is that I'm using Mint with Mesa 9.1.7, is there any chance you can test with Mesa 9.1.7 on your machine?

[tivek]

Unfortunately there are no Mesa 9.1.7 packages for Arch Linux. I did test 9.1.6 which runs the examples just fine on Ivybridge, and I also tested 9.2.0rc1 which shows the broken behavior. I tried to build Mesa from git in hopes of pinpointing the responsible commit with a bisect, but for some reason I did not manage to build any tag between 9.1.6 and 9.2.0rc1 (at least for 9.1.7 there seems to be a problem with the mapi/shared-glapi makefile).

EDIT: spoke too soon, 9.1.7 works, will try to bisect as soon as I get some sleep.

[bkaradzic]

@tivek "EDIT: spoke too soon, 9.1.7 works, will try to bisect as soon as I get some sleep."

That would be awesome! I'll try to build Mesa 10 on my machine too.

[tivek]

Found it! :) The first bad commit in the Mesa tree is 1d4dbeeaec8490a14dafefa9c647ae48b56ab952 which enables ARB_texture_multisample on Gen6+. The next step is to find Ivybridge-specific code for ARB_texture_multisample. I suspect the extension interacts in a weird way with scissor tests and this gives the buggy behavior of imgui.

[bkaradzic]

@tivek 👍

[bkaradzic]

Just tested on Mint 16 with:

OpenGL vendor string: Intel Open Source Technology Center
OpenGL renderer string: Mesa DRI Intel(R) Sandybridge Mobile
OpenGL core profile version string: 3.1 (Core Profile) Mesa 9.2.1
OpenGL core profile shading language version string: 1.40
OpenGL core profile context flags: (none)
OpenGL core profile extensions:
OpenGL version string: 3.0 Mesa 9.2.1
OpenGL shading language version string: 1.30

And bug exist.

[r-lyeh-archived]

Hey guys,

I found this issue on Mononen's original code too. I fixed it long time ago but I currently cannot remember what OpenGL command is unbalanced.

As a quick fix, you may push all attribs before rendering UI, and pop them after rendering UI.
That should work for now.

[bkaradzic]

@r-lyeh It's different because imgui here is not using GL directly. Also this issue doesn't exist with NV (noveau or propriatery drivers), OpenGL on Windows and OSX (nor earlier version of Mesa), nor OpenGL ES on Android (which uses the same code path). Definitely Mesa/Intel driver problem, I just need to figure out how to get minimal repro for them or @tivek might find exact problem with bisect.

[tivek]

Lately I wasn't able to tinker with the code as much as I would have liked. The problem I have with git bisect is that it only finds the commit to Mesa which enables an until-then unused code for ARB_texture_multisample on Ivybridge. I would have to bisect older commits while still merging the enabling commit. This could become messy (noob here :), but it's worth a shot. Any tips? Or do I just git bisect and then git cherry-pick?

As soon as I can I am going to reduce one of the bgfx examples to bare minimum that displays the problem and then use BuGLe to print out the list of GL calls. Then we'll see what to do next.