BACKENDS: OPENGL: Alpha channel changes cause color regression on PowerBook G4

[dwatteau]

I'm seeing a regression with commit 6342045cc3e23dc9f685241d89ad3da478885f70 ("BACKENDS: OPENGL: Don't take the alpha channel from the game texture") (as found by a Git bisect).

This appears with any game (but not the menu), only in opengl mode, not in surfacesdl mode.

ScummVM 2.8.x doesn't have this problem. I don't think it's OS-related, because I see this issue on both OpenBSD/macppc and Debian ppc.

The machine is a G4 PowerBook5,6 (big-endian PPC processor), with an ATI Radeon Mobility 9600/9700 (M10/M11) GPU.

I don't know if the issue is endian-related, or just old-GPU-related.

Attaching some files below. Feel free to ping me for any test!

[dwatteau]

Buggy color behavior with OpenGL renderer

[dwatteau]

Normal color behavior with surfacesdl renderer

[dwatteau]

-d6 debug output with OpenGL renderer

[dwatteau]

glxinfo output on Debian and OpenBSD

[dwatteau]

FWIW, as suggested by lephilousophe, doing this change in backends/graphics/opengl/framebuffer.cpp:

                case kBlendModeOpaque:
                        GL_CALL(glEnable(GL_BLEND));
-                       GL_CALL(glBlendColor(1.f, 1.f, 1.f, 0.f));
+                       GL_CALL(glBlendColor(1.f, 1.f, 0.f, 1.f));
                        GL_CALL(glBlendFunc(GL_CONSTANT_COLOR, GL_ONE_MINUS_CONSTANT_COLOR));
                        break;

or using glBlendColor(1.f, 1.f, 1.f, 1.f), make the color problem go away. Putting a value of 0.f for the red or green parameters doesn't fix the issue, however.

I see that both my Debian and OpenBSD systems use the same Mesa release; I'll try updating one of them to Mesa 23.x to see if it fixes anything.

[dwatteau]

As a summary of what's been said on Discord:

• this only happens on systems using Mesa; there's no problem on OSX ppc
• this indeed looks like a bug in Mesa with r300, which is not really maintained on big-endian systems anymore: <​https://gitlab.freedesktop.org/mesa/mesa/-/issues/7720>

lephilousophe made a small test program here: <​https://gist.github.com/lephilousophe/1f8bb6757a8d621e6400fc9edd57a306>. I confirm that it shows the expected behavior on x86-64, and buggy output on this machine.

Updating from Mesa 22.3.6 to Mesa 23.1.9 didn't fix the issue. Then I tried updating to Mesa 23.3.6, but that version has some other (unrelated) problems. I'll try finding some time to fix this, before I report the issue to Mesa.

Anyway, big-endian is becoming more and more of a niche thing, so more and more bugs will appear on these untested systems. It's probably time for me to freeze my PPC dev environment to something older, but more reliable.

Anyway, yes, we can consider this as a Mesa bug, not a ScummVM bug.