OGL: Stereoscopic 3D Support

[CrossVR]

This branch adds side-by-side stereoscopic 3D support through a single pass with a geometry shader. The geometry shader handles the duplication of the geometry and will project each vertex according to the projection matrix for each eye.

The stereoscopic projection matrix is generated from the normal projection by means of a shearing transformation (courtesy of @comex).

All framebuffers have been changed to layered framebuffers (courtesy of @degasus) each layer represents one eye. Using this layered rendering method means we already support EFB2Tex copies with the correct stereoscopic projection.

When testing the current build keep MSAA and XFB disabled, these options are not yet compatible with stereoscopy

To-Do

This PR

• Fix SBS pillar boxing for 4:3 ratio.
• Add options to customize stereoscopy parameters.
• Add support for top-bottom and anaglyph.
• Prevent @Parlane from ticking to-do boxes.

Future PR

• Add stereoscopy support for EFB format emulation.
• Add support for layered EFB2Tex copies with MSAA enabled.
• Add stereoscopy support to Virtual XFB.
• Add support for Quad Buffers and 3D display modes.
• Allow users to fill in their real IPD in mm.

[JMC47]

The initial support works perfectly on my crosseyed approach. Game compatibility seems identical to 3D Vision, a lot of the same errors/problems happen.

note: I had to modify the code to flip the eyes.

[CrossVR]

@JMC47 Which errors are we talking about specifically? I can use them to test compatibility.

[JMC47]

F-Zero GX has lines on the menus. Olimar's Health Bar has really weird depths in Pikmin 2, Luigi's Mansion's shadows are offset per eye. Pokemon Colosseum's Shadows are just way off, and the ground disappears at most settings, you have to turn the settings WAY down.

[comex]

Nice.

I haven't looked in detail, but are the changes in API usage here capable of causing any compatibility or performance problems?

[degasus]

@comex On bad driver, always. But in theory, layered rendering with one layer should have no overhead (tested by JMC already). There is also a bit slowdown because of the harder uid checker (could still be optimized) and the bigger vertex constant buffer. But this should both be less than 1%.

[degasus]

3D branch still works on Qualcomm

It also works fine on all of my tested plattforms but llvmpipe, but I think is a llvmpipe issue.

LGTM

[CrossVR]

@degasus @Sonicadvance1 I made some small changes to the way the stereoscopy settings are handled. It should functionally behave the same, but it will allow us to more easily tweak the stereoscopy settings.

I also did away with the Mono Depth EFB option in the settings which only has a very specific use and is just confusing to users.

[CrossVR]

On IRC @degasus and me decided it would be best to merge the PR in a disabled state for now. It would allow us to review the feature in parts and releasing it only when we feel it is completely ready.

We'd also be able to concentrate on 2D regressions before we start fixing 3D bugs.

[Icekhaos]

And here's yet another enhancement. Good job folks, keep this work up.

[Kodiack]

I don't play with stereoscopic 3D, although I do play with OpenGL. Yet, this commit still seems to have introduced a significant performance regression on my R9 290-based system. Prior to 4350 I am able to sustain a solid 41 FPS on the Skyward Sword title screen with 4xIR and 8xMSAA. After this commit, I only get 20 FPS with the same settings; I have difficulty even sustaining 30 FPS at 4xMSAA.

Again, I do not utilise stereoscopic 3D, so I am unsure why this is occurring, but it is definitely related to this pull request.

[CrossVR]

@Kodiack Could you submit this to the bug tracker with your system specifications? We were afraid this might happen, which is the reason we merged this early in a disabled state.

[Kodiack]

@Armada651 Not a problem. I'll go ahead and submit a report now. Cheers!