Implement primitive restart support

Maxwell3D also supports using an arbitrary restart index value however no games are known to use this so leave it for now.

app/src/main/cpp/skyline/gpu.cpp

@@ -126,7 +126,17 @@ namespace skyline::gpu {
                                          const vk::raii::PhysicalDevice &physicalDevice,
                                          decltype(vk::DeviceQueueCreateInfo::queueCount) &vkQueueFamilyIndex,
                                          TraitManager &traits) {
-        auto deviceFeatures2{physicalDevice.getFeatures2<vk::PhysicalDeviceFeatures2, vk::PhysicalDeviceCustomBorderColorFeaturesEXT, vk::PhysicalDeviceVertexAttributeDivisorFeaturesEXT, vk::PhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT, vk::PhysicalDeviceShaderFloat16Int8Features, vk::PhysicalDeviceShaderAtomicInt64Features, vk::PhysicalDeviceUniformBufferStandardLayoutFeatures, vk::PhysicalDeviceShaderDrawParametersFeatures, vk::PhysicalDeviceProvokingVertexFeaturesEXT>()};
+        auto deviceFeatures2{physicalDevice.getFeatures2<
+            vk::PhysicalDeviceFeatures2,
+            vk::PhysicalDeviceCustomBorderColorFeaturesEXT,
+            vk::PhysicalDeviceVertexAttributeDivisorFeaturesEXT,
+            vk::PhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT,
+            vk::PhysicalDeviceShaderFloat16Int8Features,
+            vk::PhysicalDeviceShaderAtomicInt64Features,
+            vk::PhysicalDeviceUniformBufferStandardLayoutFeatures,
+            vk::PhysicalDeviceShaderDrawParametersFeatures,
+            vk::PhysicalDeviceProvokingVertexFeaturesEXT,
+            vk::PhysicalDevicePrimitiveTopologyListRestartFeaturesEXT>()};
         decltype(deviceFeatures2) enabledFeatures2{}; // We only want to enable features we required due to potential overhead from unused features
 
         #define FEAT_REQ(structName, feature)                                            \
@@ -156,7 +166,11 @@ namespace skyline::gpu {
                 throw exception("Cannot find Vulkan device extension: \"{}\"", requiredExtension.data());
         }
 
-        auto deviceProperties2{physicalDevice.getProperties2<vk::PhysicalDeviceProperties2, vk::PhysicalDeviceDriverProperties, vk::PhysicalDeviceFloatControlsProperties, vk::PhysicalDeviceSubgroupProperties>()};
+        auto deviceProperties2{physicalDevice.getProperties2<
+            vk::PhysicalDeviceProperties2,
+            vk::PhysicalDeviceDriverProperties,
+            vk::PhysicalDeviceFloatControlsProperties,
+            vk::PhysicalDeviceSubgroupProperties>()};
 
         traits = TraitManager(deviceFeatures2, enabledFeatures2, deviceExtensions, enabledExtensions, deviceProperties2);
         traits.ApplyDriverPatches(context);

app/src/main/cpp/skyline/gpu/interconnect/graphics_context.h

@@ -1743,6 +1743,33 @@ namespace skyline::gpu::interconnect {
       private:
         vk::PipelineInputAssemblyStateCreateInfo inputAssemblyState{};
 
+        void ValidatePrimitiveRestartState() {
+            if (inputAssemblyState.primitiveRestartEnable) {
+                switch (inputAssemblyState.topology) {
+                    case vk::PrimitiveTopology::eLineStrip:
+                    case vk::PrimitiveTopology::eTriangleStrip:
+                    case vk::PrimitiveTopology::eTriangleFan:
+                    case vk::PrimitiveTopology::eLineStripWithAdjacency:
+                    case vk::PrimitiveTopology::eTriangleStripWithAdjacency:
+                        break; // Doesn't need any extension
+
+                    case vk::PrimitiveTopology::ePointList:
+                    case vk::PrimitiveTopology::eLineList:
+                    case vk::PrimitiveTopology::eTriangleList:
+                    case vk::PrimitiveTopology::eLineListWithAdjacency:
+                    case vk::PrimitiveTopology::eTriangleListWithAdjacency:
+                        if (!gpu.traits.supportsTopologyListRestart)
+                            Logger::Warn("GPU doesn't support primitive restart with list topologies");
+                        break;
+
+                    case vk::PrimitiveTopology::ePatchList:
+                        if (!gpu.traits.supportsTopologyPatchListRestart)
+                            Logger::Warn("GPU doesn't support primitive restart with patch list topology");
+                        break;
+                }
+            }
+        }
+
       public:
         void SetPrimitiveTopology(maxwell3d::PrimitiveTopology topology) {
             auto[vkTopology, shaderTopology] = [topology]() -> std::tuple<vk::PrimitiveTopology, ShaderCompiler::InputTopology> {
@@ -1778,6 +1805,10 @@ namespace skyline::gpu::interconnect {
             UpdateRuntimeInformation(runtimeInfo.input_topology, shaderTopology, maxwell3d::PipelineStage::Geometry);
         }
 
+        void SetPrimitiveRestartEnabled(bool enable) {
+            inputAssemblyState.primitiveRestartEnable = enable;
+        }
+
         /* Index Buffer */
       private:
         struct IndexBuffer {
@@ -2540,6 +2571,8 @@ namespace skyline::gpu::interconnect {
       public:
         template<bool IsIndexed>
         void Draw(u32 count, u32 first, i32 vertexOffset = 0) {
+            ValidatePrimitiveRestartState();
+
             // Shader + Binding Setup
             auto programState{CompileShaderProgramState()};
             auto descriptorSet{gpu.descriptor.AllocateSet(*programState.descriptorSetLayout)};

app/src/main/cpp/skyline/gpu/trait_manager.cpp

@@ -6,7 +6,7 @@
 
 namespace skyline::gpu {
     TraitManager::TraitManager(const DeviceFeatures2 &deviceFeatures2, DeviceFeatures2 &enabledFeatures2, const std::vector<vk::ExtensionProperties> &deviceExtensions, std::vector<std::array<char, VK_MAX_EXTENSION_NAME_SIZE>> &enabledExtensions, const DeviceProperties2 &deviceProperties2) : quirks(deviceProperties2.get<vk::PhysicalDeviceProperties2>().properties, deviceProperties2.get<vk::PhysicalDeviceDriverProperties>()) {
-        bool hasCustomBorderColorExt{}, hasShaderAtomicInt64Ext{}, hasShaderFloat16Int8Ext{}, hasShaderDemoteToHelperExt, hasVertexAttributeDivisorExt, hasProvokingVertexExt{};
+        bool hasCustomBorderColorExt{}, hasShaderAtomicInt64Ext{}, hasShaderFloat16Int8Ext{}, hasShaderDemoteToHelperExt{}, hasVertexAttributeDivisorExt{}, hasProvokingVertexExt{}, hasPrimitiveTopologyListRestartExt{};
         bool supportsUniformBufferStandardLayout{}; // We require VK_KHR_uniform_buffer_standard_layout but assume it is implicitly supported even when not present
 
         for (auto &extension : deviceExtensions) {
@@ -42,6 +42,7 @@ namespace skyline::gpu {
                 EXT_SET("VK_KHR_shader_float16_int8", hasShaderFloat16Int8Ext);
                 EXT_SET("VK_KHR_shader_float_controls", supportsFloatControls);
                 EXT_SET("VK_KHR_uniform_buffer_standard_layout", supportsUniformBufferStandardLayout);
+                EXT_SET("VK_EXT_primitive_topology_list_restart", hasPrimitiveTopologyListRestartExt);
             }
 
             #undef EXT_SET
@@ -110,6 +111,13 @@ namespace skyline::gpu {
             Logger::Warn("Cannot find VK_KHR_uniform_buffer_standard_layout, assuming implicit support");
         }
 
+        if (hasPrimitiveTopologyListRestartExt) {
+            FEAT_SET(vk::PhysicalDevicePrimitiveTopologyListRestartFeaturesEXT, primitiveTopologyListRestart, supportsTopologyListRestart)
+            FEAT_SET(vk::PhysicalDevicePrimitiveTopologyListRestartFeaturesEXT, primitiveTopologyPatchListRestart, supportsTopologyPatchListRestart)
+        } else {
+            enabledFeatures2.unlink<vk::PhysicalDevicePrimitiveTopologyListRestartFeaturesEXT>();
+        }
+
         #undef FEAT_SET
 
         if (supportsFloatControls)
@@ -122,8 +130,8 @@ namespace skyline::gpu {
 
     std::string TraitManager::Summary() {
         return fmt::format(
-            "\n* Supports U8 Indices: {}\n* Supports Sampler Mirror Clamp To Edge: {}\n* Supports Sampler Reduction Mode: {}\n* Supports Custom Border Color (Without Format): {}\n* Supports Last Provoking Vertex: {}\n* Supports Logical Operations: {}\n* Supports Vertex Attribute Divisor: {}\n* Supports Vertex Attribute Zero Divisor: {}\n* Supports Multiple Viewports: {}\n* Supports Shader Viewport Index: {}\n* Supports SPIR-V 1.4: {}\n* Supports Shader Invocation Demotion: {}\n* Supports 16-bit FP: {}\n* Supports 8-bit Integers: {}\n* Supports 16-bit Integers: {}\n* Supports 64-bit Integers: {}\n* Supports Atomic 64-bit Integers: {}\n* Supports Floating Point Behavior Control: {}\n* Supports Image Read Without Format: {}\n* Supports Subgroup Vote: {}\n* Subgroup Size: {}",
+            "\n* Supports U8 Indices: {}\n* Supports Sampler Mirror Clamp To Edge: {}\n* Supports Sampler Reduction Mode: {}\n* Supports Custom Border Color (Without Format): {}\n* Supports Last Provoking Vertex: {}\n* Supports Logical Operations: {}\n* Supports Vertex Attribute Divisor: {}\n* Supports Vertex Attribute Zero Divisor: {}\n* Supports Multiple Viewports: {}\n* Supports Shader Viewport Index: {}\n* Supports SPIR-V 1.4: {}\n* Supports Shader Invocation Demotion: {}\n* Supports 16-bit FP: {}\n* Supports 8-bit Integers: {}\n* Supports 16-bit Integers: {}\n* Supports 64-bit Integers: {}\n* Supports Atomic 64-bit Integers: {}\n* Supports Floating Point Behavior Control: {}\n* Supports Image Read Without Format: {}\n* Supports List Primitive Topology Restart: {}\n* Supports Patch List Primitive Topology Restart: {}\n* Supports Subgroup Vote: {}\n* Subgroup Size: {}",
-            supportsUint8Indices, supportsSamplerMirrorClampToEdge, supportsSamplerReductionMode, supportsCustomBorderColor, supportsLastProvokingVertex, supportsLogicOp, supportsVertexAttributeDivisor, supportsVertexAttributeZeroDivisor, supportsMultipleViewports, supportsShaderViewportIndexLayer, supportsSpirv14, supportsShaderDemoteToHelper, supportsFloat16, supportsInt8, supportsInt16, supportsInt64, supportsAtomicInt64, supportsFloatControls, supportsImageReadWithoutFormat, supportsSubgroupVote, subgroupSize
+            supportsUint8Indices, supportsSamplerMirrorClampToEdge, supportsSamplerReductionMode, supportsCustomBorderColor, supportsLastProvokingVertex, supportsLogicOp, supportsVertexAttributeDivisor, supportsVertexAttributeZeroDivisor, supportsMultipleViewports, supportsShaderViewportIndexLayer, supportsSpirv14, supportsShaderDemoteToHelper, supportsFloat16, supportsInt8, supportsInt16, supportsInt64, supportsAtomicInt64, supportsFloatControls, supportsImageReadWithoutFormat, supportsTopologyListRestart, supportsTopologyPatchListRestart, supportsSubgroupVote, subgroupSize
         );
     }
 

app/src/main/cpp/skyline/gpu/trait_manager.h

@@ -32,6 +32,8 @@ namespace skyline::gpu {
         bool supportsFloatControls{}; //!< If extensive control over FP behavior is exposed (with VK_KHR_shader_float_controls)
         vk::PhysicalDeviceFloatControlsProperties floatControls{}; //!< Specifics of FP behavior control (All members will be zero'd out when unavailable)
         bool supportsImageReadWithoutFormat{}; //!< If a storage image can be read without a format
+        bool supportsTopologyListRestart{}; //!< If the device supports using primitive restart for topology lists (with VK_EXT_primitive_topology_list_restart)
+        bool supportsTopologyPatchListRestart{}; //!< If the device supports using primitive restart for topology patch lists (with VK_EXT_primitive_topology_list_restart)
         bool supportsSubgroupVote{}; //!< If subgroup votes are supported in shaders with SPV_KHR_subgroup_vote
         u32 subgroupSize{}; //!< Size of a subgroup on the host GPU
 
@@ -56,9 +58,23 @@ namespace skyline::gpu {
 
         TraitManager() = default;
 
-        using DeviceProperties2 = vk::StructureChain<vk::PhysicalDeviceProperties2, vk::PhysicalDeviceDriverProperties, vk::PhysicalDeviceFloatControlsProperties, vk::PhysicalDeviceSubgroupProperties>;
+        using DeviceProperties2 = vk::StructureChain<
+            vk::PhysicalDeviceProperties2,
+            vk::PhysicalDeviceDriverProperties,
+            vk::PhysicalDeviceFloatControlsProperties,
+            vk::PhysicalDeviceSubgroupProperties>;
 
-        using DeviceFeatures2 = vk::StructureChain<vk::PhysicalDeviceFeatures2, vk::PhysicalDeviceCustomBorderColorFeaturesEXT, vk::PhysicalDeviceVertexAttributeDivisorFeaturesEXT, vk::PhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT, vk::PhysicalDeviceShaderFloat16Int8Features, vk::PhysicalDeviceShaderAtomicInt64Features, vk::PhysicalDeviceUniformBufferStandardLayoutFeatures, vk::PhysicalDeviceShaderDrawParametersFeatures, vk::PhysicalDeviceProvokingVertexFeaturesEXT>;
+        using DeviceFeatures2 = vk::StructureChain<
+            vk::PhysicalDeviceFeatures2,
+            vk::PhysicalDeviceCustomBorderColorFeaturesEXT,
+            vk::PhysicalDeviceVertexAttributeDivisorFeaturesEXT,
+            vk::PhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT,
+            vk::PhysicalDeviceShaderFloat16Int8Features,
+            vk::PhysicalDeviceShaderAtomicInt64Features,
+            vk::PhysicalDeviceUniformBufferStandardLayoutFeatures,
+            vk::PhysicalDeviceShaderDrawParametersFeatures,
+            vk::PhysicalDeviceProvokingVertexFeaturesEXT,
+            vk::PhysicalDevicePrimitiveTopologyListRestartFeaturesEXT>;
 
         TraitManager(const DeviceFeatures2 &deviceFeatures2, DeviceFeatures2 &enabledFeatures2, const std::vector<vk::ExtensionProperties> &deviceExtensions, std::vector<std::array<char, VK_MAX_EXTENSION_NAME_SIZE>> &enabledExtensions, const DeviceProperties2 &deviceProperties2);
 

app/src/main/cpp/skyline/soc/gm20b/engines/maxwell_3d.cpp

@@ -494,6 +494,10 @@ namespace skyline::soc::gm20b::engine::maxwell3d {
                     context.SetPrimitiveTopology(vertexBeginGl.topology);
                 })
 
+                ENGINE_CASE(primitiveRestartEnable, {
+                    context.SetPrimitiveRestartEnabled(primitiveRestartEnable);
+                })
+
                 ENGINE_STRUCT_CASE(constantBufferSelector, size, {
                     context.SetConstantBufferSelectorSize(size);
                 })

app/src/main/cpp/skyline/soc/gm20b/engines/maxwell_3d.h

@@ -227,6 +227,9 @@ namespace skyline::soc::gm20b::engine::maxwell3d {
             Register<0x585, u32> vertexEndGl; //!< Method-only register with no real value, used after calling vertexBeginGl to invoke the draw
             Register<0x586, type::VertexBeginGl> vertexBeginGl; //!< Similar to glVertexBegin semantically, supplies a primitive topology for draws alongside instancing data
 
+            Register<0x591, u32> primitiveRestartEnable;
+            Register<0x592, u32> primitiveRestartIndex;
+
             Register<0x5A1, u32> provokingVertexIsLast;
 
             Register<0x5F2, type::IndexBuffer> indexBuffer;