OpenGL 1/2 context parameters aren't always passed correctly on Linux

[gravgun]

Hello everyone,
when I try to apply antialias on a sf::ContextSettings and pass it to sf::RenderWindow, it doesn't work; i just get a default OpenGL 1.x/2.x context.
I don't think it is the expected behavior, how do I get to enable antialias then?

The culprit is src/SFML/Window/Linux/GlxContext.cpp, where an test is being done; introduced by Laurent when he fixed bug #258:

// Create the OpenGL context -- first try context versions >= 3.0 if it is requested (they require special code)
if (m_settings.majorVersion >= 3)
{

and the initialization code passes everything to glXCreateContextAttribsARB, but this test restricts all of this to OpenGL 3+.
And what does OpenGL 1.x/2.x get? Nothing. Basically I removed this test (and && (m_settings.majorVersion >= 3) too), and the whole thing worked: initialization for OpenGL 1.x/2.x is done the same way as 3+...

Could this get fixed please?

[Laurent]

The 2.x context creation code is supposed to work too. The visual is chosen with GlxContext::selectBestVisual, which correctly takes multisampling in account.

[gravgun]

The problem is that the way it selects the best visual is not suited here, example:
GlContext::evaluateFormat check for best fitting context, but can produce unwanted output; and my Intel chipset can do antialias, so that's not the problem.
Good case:
Want: 32bpp, 16 bits depth, 8 bits stencil, x4 antialias
Have: 32bpp, 16 bits depth, 8 bits stencil, x4 antialias
      0   +  0      +       0       +        0 = 0, best, everything is OK

Bad case:
Want: 32bpp, 8 bits depth, 8 bits stencil, x2 antialias
Have: 32bpp, 8 bits depth, 8 bits stencil, x0 antialias
      0   +  0      +       0       +       2 = 2, gets chosen but I want antialias
      32bpp, 16 bits depth, 8 bits stencil, x2 antialias
      0   +  8      +       0       +        0 = 8
Even if values seem illogical to you, I seem to be in the bad case (I haven't got the actual values though).
This behavior is due to the std::abses in GlContext::evaluateFormat, that makes the rating worse if the avaliable settings is actually better; which is both a bad and good thing: it gets the closest context settings, but better isn't necessarily worse.

As I said, using the OpenGL 3/4 init code does work for 1.x/2.x; and even if it fails the fallback will still be available to get the closest settings possible. So I think it'd be better trying to create a context with exact options then falling back, regardless of the requested OpenGL version.

[Laurent]

GlContext::evaluateFormat check for best fitting context, but can produce unwanted output

Yes, the selection algorithm is not perfect. I guess there will always be corner cases that will produce unexpected results.

As I said, using the OpenGL 3/4 init code does work for 1.x/2.x; and even if it fails the fallback will still be available to get the closest settings possible. So I think it'd be better trying to create a context with exact options then falling back, regardless of the requested OpenGL version.

Looks like a good suggestion

Thanks for your feedback, it is very helpful.