Expose experimental Vulkan features using internal Metal API

[spnda]

As noted by @cdavis5e in #1657, there's a few things in Metal which are not documented or exposed. These features could possibly be used to experiment supporting some Vulkan features that would not be supportable otherwise. The initial idea is to have a compile-time option which would enable these experimental features, as using these private APIs would not allow shipping an app with MoltenVK into the App Store and should therefore not be the default. This issue shall act as some sort of a tracker of potential features and their implementations by looking into the symbols and Objective-C metadata.

5 patches have already been written to support some features (although not entirely stable): https://github.com/KhronosGroup/MoltenVK/issues/1657#issuecomment-1197438815

I have also just gone ahead and extracted the symbols of the Metal binary found in the Metal.framework shipped with the latest macOS 13 beta, meaning with support for Metal 3 and Metal IO. I couldn't get class-dump to not crash with this binary, so this is just the basic output of nm: metal_symbols.txt metalkit_symbols.txt

Some few things I've found in metal_symbols.txt:

• -[MTLHeap _gpuAddress]
• MTLDeviceFeatureQueries contains A LOT of interesting queries about GPU features, of which most are not exposed normally.
• Metal acceleration structures seem to be built using MTLGPUBVHBuilder, which seems to be more like something from MPS.
• _MTLDevice has a lot of other interesting queries about GPU capabilities.
• _MTLPipelineCache is a thing.
• setLineWidth: (as used by one of the patches linked above) is not listed anymore.
• -[MTLRenderPipelineColorAttachmentDescriptor setLogicOpEnabled:] is not listed anymore. However, -[MTLRenderPipelineDescriptor setLogicOperation:] is available.
• -[MTLToolsRenderCommandEncoder useResourceGroup:usage:stages:]... I wonder what resource groups are. (found in MetalTools)

[billhollings]

setLineWidth: (as used by one of the patches linked above) is not listed anymore.

Nice sleuthing.

Interestingly, -[_MTLDevice maxLineWidth] is still there.

Hmmm...unless it's moved to a different library that is accessible...this is the kind of thing where it can get really messy to try to support these kind of private-API calls...and frustrating for users if something arrives for a while and then gets taken away.

[TellowKrinkle]

Just a heads up: Just because the method is there, doesn't mean it works, or works well.

Some others I've found:

• Intel Skylake GPUs and up support framebuffer fetch, but the GPU locks up if you try to use both that and dual source blending at the same time

• MTLRenderPipelineDescriptor has things for logic ops, but I've only gotten it to work on Intel Skylake+ (AMD, Nvidia, and older Intel act like blending is disabled when you set those)

  Test script and results from an AMD and Intel GPU

  ```objc #import #import typedef NS_ENUM(NSUInteger, MTLLogicOperation) { MTLLogicOperationClear = 0, MTLLogicOperationSet = 1, MTLLogicOperationCopy = 2, MTLLogicOperationCopyInverted = 3, MTLLogicOperationNoOp = 4, MTLLogicOperationInvert = 5, MTLLogicOperationAnd = 6, MTLLogicOperationNand = 7, MTLLogicOperationOr = 8, MTLLogicOperationNor = 9, MTLLogicOperationXor = 10, MTLLogicOperationEquiv = 11, MTLLogicOperationAndReverse = 12, MTLLogicOperationAndInverted = 13, MTLLogicOperationOrReverse = 14, MTLLogicOperationOrInverted = 15, }; @interface MTLRenderPipelineDescriptor() - (void)setLogicOperationEnabled:(BOOL)enabled; - (BOOL)isLogicOperationEnabled; - (MTLLogicOperation)logicOperation; - (void)setLogicOperation:(MTLLogicOperation)operation; @end NSString* libstring = @"" "#include \n" "using namespace metal;\n" "\n" "vertex float4 fullscreen_triangle(uint vid [[vertex_id]]) {\n" " return float4(vid & 1 ? 3 : -1, vid & 2 ? 3 : -1, 0, 1);\n" "}\n" "\n" "fragment float4 fs_base() {\n" " return 0b1100 / 255.f;\n" "}\n" "fragment float4 fs_logic() {\n" " return 0b1010 / 255.f;\n" "}\n"; int main(int argc, const char * argv[]) { @autoreleasepool { for (id gpu in MTLCopyAllDevices()) { id lib = [gpu newLibraryWithSource:libstring options:nil error:nil]; printf("Testing %s\n", [[gpu name] UTF8String]); MTLRenderPipelineDescriptor* desc = [MTLRenderPipelineDescriptor new]; [desc setVertexFunction:[lib newFunctionWithName:@"fullscreen_triangle"]]; [desc setFragmentFunction:[lib newFunctionWithName:@"fs_base"]]; [desc.colorAttachments[0] setPixelFormat:MTLPixelFormatRGBA8Unorm]; NSError* err = nil; id pipeBase = [gpu newRenderPipelineStateWithDescriptor:desc error:&err]; MTLTextureDescriptor* tdesc = [MTLTextureDescriptor texture2DDescriptorWithPixelFormat:MTLPixelFormatRGBA8Unorm width:16 height:1 mipmapped:NO]; [tdesc setUsage:MTLTextureUsageRenderTarget]; [tdesc setStorageMode:MTLStorageModePrivate]; id tex = [gpu newTextureWithDescriptor:tdesc]; id buf = [gpu newBufferWithLength:4 options:MTLResourceStorageModeShared]; MTLRenderPassDescriptor* rpdesc = [MTLRenderPassDescriptor new]; id q = [gpu newCommandQueue]; [desc setLogicOperationEnabled:YES]; [desc setFragmentFunction:[lib newFunctionWithName:@"fs_logic"]]; for (int i = 0; i < 16; i++) { @autoreleasepool { rpdesc.colorAttachments[0].texture = tex; rpdesc.colorAttachments[0].loadAction = MTLLoadActionLoad; rpdesc.colorAttachments[0].storeAction = MTLStoreActionStore; [desc setLogicOperation:i]; id pipeLogic = [gpu newRenderPipelineStateWithDescriptor:desc error:&err]; if (err) { printf("Failed to make logic op pipe: %s\n", [[err localizedDescription] UTF8String]); } id cb = [q commandBuffer]; id enc = [cb renderCommandEncoderWithDescriptor:rpdesc]; [enc setRenderPipelineState:pipeBase]; [enc drawPrimitives:MTLPrimitiveTypeTriangle vertexStart:0 vertexCount:3]; [enc setRenderPipelineState:pipeLogic]; [enc drawPrimitives:MTLPrimitiveTypeTriangle vertexStart:0 vertexCount:3]; [enc endEncoding]; id download = [cb blitCommandEncoder]; [download copyFromTexture:tex sourceSlice:0 sourceLevel:0 sourceOrigin:MTLOriginMake(0, 0, 0) sourceSize:MTLSizeMake(1, 1, 1) toBuffer:buf destinationOffset:0 destinationBytesPerRow:4 destinationBytesPerImage:4]; [download endEncoding]; [cb commit]; [cb waitUntilCompleted]; unsigned int result = *(unsigned int*)[buf contents]; printf("1100(D) 1010(S) → %c%c%c%c\n", i, result & 8 ? '1' : '0', result & 4 ? '1' : '0', result & 2 ? '1' : '0', result & 1 ? '1' : '0'); } } } } return 0; } ``` Results: ``` Testing Intel(R) UHD Graphics 630 1100(D) 1010(S) → 0000 1100(D) 1010(S) → 1111 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 0101 1100(D) 1010(S) → 1100 1100(D) 1010(S) → 0011 1100(D) 1010(S) → 1000 1100(D) 1010(S) → 0111 1100(D) 1010(S) → 1110 1100(D) 1010(S) → 0001 1100(D) 1010(S) → 0110 1100(D) 1010(S) → 1001 1100(D) 1010(S) → 0010 1100(D) 1010(S) → 0100 1100(D) 1010(S) → 1011 1100(D) 1010(S) → 1101 Testing AMD Radeon Pro 5600M 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 1100(D) 1010(S) → 1010 ```

I'd expect setLineWidth: to continue to exist as long as OpenGL does, at least on Apple GPUs, since I'm pretty sure they use it internally for the OpenGL → Metal translation layer, which is used for OpenGL support on Apple. I think it's not showing up there because protocols aren't emitted into binaries, and no classes implementing MTLRenderCommandEncoder are in Metal.framework (they're all in the individual drivers).

The stand-in command encoder used by the Metal frame capture API (e.g. if you run from Xcode) implements more methods than the underlying command encoder may, and will blindly forward them. So just checking if the command encoder has a method isn't enough, you have to call it from a @try {} to check if it throws an exception (or detect capture and api validation encoders and get their underlying encoders). (e.g. capture encoders have a setPointSize: in addition to the setLineWidth:, which seems to actually exist on Apple GPUs but nothing else)

If you really wanted to expose them in MoltenVK, you'd have to either individually whitelist/blacklist GPUs, or test each feature at runtime to ensure they actually work.

[italomandara]

Just a heads up: Just because the method is there, doesn't mean it works, or works well.

Some others I've found:

• Intel Skylake GPUs and up support framebuffer fetch, but the GPU locks up if you try to use both that and dual source blending at the same time

• MTLRenderPipelineDescriptor has things for logic ops, but I've only gotten it to work on Intel Skylake+ (AMD, Nvidia, and older Intel act like blending is disabled when you set those)

  Test script and results from an AMD and Intel GPU ...

I'd expect setLineWidth: to continue to exist as long as OpenGL does, at least on Apple GPUs, since I'm pretty sure they use it internally for the OpenGL → Metal translation layer, which is used for OpenGL support on Apple. I think it's not showing up there because protocols aren't emitted into binaries, and no classes implementing MTLRenderCommandEncoder are in Metal.framework (they're all in the individual drivers).

The stand-in command encoder used by the Metal frame capture API (e.g. if you run from Xcode) implements more methods than the underlying command encoder may, and will blindly forward them. So just checking if the command encoder has a method isn't enough, you have to call it from a @try {} to check if it throws an exception (or detect capture and api validation encoders and get their underlying encoders). (e.g. capture encoders have a setPointSize: in addition to the setLineWidth:, which seems to actually exist on Apple GPUs but nothing else)

If you really wanted to expose them in MoltenVK, you'd have to either individually whitelist/blacklist GPUs, or test each feature at runtime to ensure they actually work.

and maybe polyfill to "wideline -> triangles" emulation where missing?

[billhollings]

PR #2137 fixes this for wide lines.