GS: Default to OpenGL for RDNA 2 and 3 on windows.

[lightningterror]

Description of Changes

When someone buys an RDNA 3 GPU please debug the driver timeouts on windows to figure out what's causing the crashes on Vulkan, seems to be related to barriers.

RDNA 2 and higher support feature level 12.2 so use that to detect RDNA 3.

Rationale behind Changes

Avoid driver timeouts on RDNA3 on Vulkan.
We need someone with an RDNA3 GPU to debug and fix the issue which we don't have.

Suggested Testing Steps

Test to make sure rdna 2 and 3 default to gl, older vulkan capable to vulkan, legacy to dx11.

[TellowKrinkle]

Why are we including RDNA2 in this?

[RinMaru]

Why are we including RDNA2 in this?

Exactly what im thinking. I haven't had any issues on my 6700XT RDNA2 and I use VK exclusively
Windows 11 24H2
Adrenalin 24.12.1

[lightningterror]

Because otherwise we'd have to hardcode every rdna3 and possibly future gpu to default to GL instead and I don't like that, you can still switch from Auto to Vulkan for rdna2.

[mirh]

Couldn't you use some other architectural feature?
Feature level alone has the little problem that it would probably always be 12_1 on W10.

[JordanTheToaster]

Couldn't you use some other architectural feature? Feature level alone has the little problem that it would probably always be 12_1 on W10.

Both RDNA 2 and 3 are identical feature level wise you cannot use that to pick one or the other.

[AmyRoxwell]

Probably would be a good idea to make an issue page about this, just in the case an external dev with a RDNA3 GPU sees it and might be able to debug it or just for the sake to keep track on the issue.

[mirh]

Both RDNA 2 and 3 are identical feature level wise you cannot use that to pick one or the other.

In my link, shader model 6.8 is specifically singled out.
Not sure if even that could be queried from inside W10, but it seems impossible that over hundreds of caps they would all be the same.

Speaking of which, I guess resource binding could also be used to distinguish between fermi, kepler and maxwell.

[mirh]

Ok so, I checked 6700 XT vs 7700S (and also a 780M) with dx caps viewer.. And indeed aside of reading tea leaves and then hoping for the best guess I couldn't spot night and day differences.

Thankfully I then found D3d12info and D3d12infoDB and I figured I had been pretty stupid because the last compiled dxcaps exe is from before RDNA3 was even released (i.e. it doesn't have any of the fanciest new stuff).
AFAICT work graphs are the one feature which is RDNA3-only (and as also discussed by the B3D guys, it should go hand-in-hand with D3D12_SHADER_MODEL_6_8 too).

[TellowKrinkle]

There's a number of differences in Vulkan caps as well, though if you go that route, it's probably easiest to just check VK_AMD_shader_core_properties

[mirh]

I don't think so, that seems pretty much identical too (besides model-specific variations like compute units per array or total).
There are other differences here and there though, this for instance holds for openGL too (and it's the only one there):

	RDNA 2	RDNA 3
MAX_GEOMETRY_OUTPUT_VERTICES	1023	256
MAX_GEOMETRY_TOTAL_OUTPUT_COMPONENTS	4095	1024

Conversely indeed there are still many smaller ones in vulkan (including different supported extensions and sub-properties).
But last time I checked, you don't want to use non-native apis for detection if you can (both because you load a dependency that you may not need, but also because if you are trying to avoid it in the first place it's not the brightest idea on the shed).

[TellowKrinkle]

I don't think so, that seems pretty much identical too (besides model-specific variations like compute units per array or total).

I was looking at VGPRs per SIMD but it looks like AMD only raised that for the bigger RDNA3 cards, and not the smaller ones.