VideoBackends: Add Metal backend #10754
Conversation
|
Dolphin's build infrastructure won't build versions that contain merge conflicts; please rebase. |
| vendor = DriverDetails::VENDOR_APPLE; | ||
| double version = 0; | ||
| if (@available(macOS 13, iOS 16, *)) | ||
| version = 3.0; |
There was a problem hiding this comment.
Perhaps it would be better to use macOS version instead of the Metal version for DriverDetails, since Apple may fix driver bugs in minor macOS updates.
There was a problem hiding this comment.
Then wouldn't we have to figure out some versioning system that unifies iOS and macOS versions? Then again if they did fix a bug in a minor update, there's no guarantee it would come to iOS and macOS at the same time...
There was a problem hiding this comment.
Technically, you don't need to worry about iOS since Dolphin has no official iOS support. (That being said, having more things upstreamed would make my fork easier to maintain...)
There was a problem hiding this comment.
OK, so I had a chat with @JMC47 and they said it should be fine to merge some iOS support upstream.
I think we should use the OS version instead of the Metal version for DriverDetails, and just have separate entries for iOS and macOS. That's probably the easiest and cleanest way.
There was a problem hiding this comment.
Okay
I put it in as major * 100 + minor, but if you prefer we can enjoy our floating point rounding fun and go for major + minor / 100
| bool supports_apple4 = false; | ||
| if (@available(iOS 13, *)) | ||
| { | ||
| supports_mac1 = [device supportsFamily:MTLGPUFamilyMac1]; |
There was a problem hiding this comment.
You're checking for a Mac GPU family in code that runs on iOS. Is this correct?
There was a problem hiding this comment.
I thought M1 iPads might support it? Don't have one to test though.
TellowKrinkle
force-pushed
the
Metal
branch
3 times, most recently
from
caa1a84
to
dbf443a
Compare
|
@TellowKrinkle you can run |
|
Cool! My quick Wind Waker test shows quite a performance boost. However, anything containing shaders makes Dolphin crash on shader pre-compilation. My machine: |
https://github.com/KhronosGroup/SPIRV-Cross/blob/master/spirv_cross_error_handling.hpp#L49 seems debug build should print something useful |
Yes, a debug build for macOS would be a good idea… |
I did that and accepted the reasonable looking changes. The rest are on my list of things to mess with so clang-format accepts them. If you happen to know why it wants to misalign these method calls, that would be very helpful to know.
I've been having a lot of issues with uid caches not getting invalidated. We'll figure it out by the time things get merged, but for now, try deleting all the uidcache files in your |
I don't know, but some quick playing around shows that using Webkit or Mozilla tbh I personally just let the editor autoformat and ignore the specifics. |
|
Okay looks like it was an issue with the indentation level. Using longer variable names fixes it. The rest of the formatting is stuff that's fixed in the (separate) PRs for some of the other pieces of this, so it should clear up when those get merged and I rebase over |
That did the trick, thanks. Didn't find any graphical corruption, yet and it seems very stable so far. Most slowdowns I had with MoltenVK are practically gone. I can even go to higher resolutions, now, which was never possible with either OpenGL or Vulkan. Apparently, the MoltenVK overhead was indeed quite substantial. |
TellowKrinkle
force-pushed
the
Metal
branch
7 times, most recently
from
ca4d758
to
250385f
Compare
|
Rebuilt today with your latest changes and while playing PSO tested the new MSAA support and there are small line artifacts in some textures in the backgrounds with it enabled, they display correctly if you deactivate MSAA, tested with both MSAA 2x and 4x, none of this happens with the SSAA's Is not too noticeable (but its still there) if you upscale beyond 3x or 4x or 5x + enabling MSAA With MSAA Without MSAA |
|
Upon further inspection I can confirm MSAA is working as intended, the artifacts seem in my previous post were caused by the way the game handles the textures so it can be considered a normal behavior. |
Does enabling vertex rounding fix it? |
|
Vertex Rounding was already enabled when the screenshots were taken. |
|
Not sure this is accurate but this is interesting. Running 4x native and it’s consuming about 2 watts. Vulkan used to average about 14w. pCzwOZ1.-.Imgur.mp4And the power usage does go up as expected when doing other things. If this is reporting correctly, this is incredible. |
|
Yes thats how its meant to be, it is a sight to behold! |
Correcting this bit. I just check with Vulkan on latest master and it’s only using 1w more on average for the same intro. I understand it’s not a particularly demanding track but Dolphin Vulkan never consumed so little power until now. The demanding tracks definitely pushes the CPU harder and that’s where Metal shines here. |
TellowKrinkle
force-pushed
the
Metal
branch
2 times, most recently
from
1a463c3
to
9a93ce0
Compare
TellowKrinkle
force-pushed
the
Metal
branch
2 times, most recently
from
c5659dd
to
2d70626
Compare
Metal pipelines hold less stuff than Dolphin pipelines, so duplicates will appear
Not worth the extra code
|
Looks like this has been reviewed by a macOS maintainer and a GPU backend maintainer. Only one thing left to do. |
Adds a Metal backend.
Currently includes code from lots of other PRs. Will shrink once those merge. PRing now because people wanted builds to test.