Implement QuadList support for non-indexed draws

app/CMakeLists.txt

@@ -171,6 +171,7 @@ add_library(skyline SHARED
         ${source_DIR}/skyline/gpu/shader_manager.cpp
         ${source_DIR}/skyline/gpu/interconnect/command_executor.cpp
         ${source_DIR}/skyline/gpu/interconnect/command_nodes.cpp
+        ${source_DIR}/skyline/gpu/interconnect/conversion/quads.cpp
         ${source_DIR}/skyline/soc/smmu.cpp
         ${source_DIR}/skyline/soc/host1x/syncpoint.cpp
         ${source_DIR}/skyline/soc/host1x/command_fifo.cpp

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

@@ -87,6 +87,16 @@ namespace skyline::gpu {
             return backing.vkBuffer;
         }
 
+        /**
+         * @return A span over the backing of this buffer
+         * @note This operation **must** be performed only on host-only buffers since synchronization is handled internally for guest-backed buffers
+         */
+        span<u8> GetBackingSpan() {
+            if (guest)
+                throw exception("Attempted to get a span of a guest-backed buffer");
+            return span<u8>(backing);
+        }
+
         Buffer(GPU &gpu, GuestBuffer guest);
 
         /**

app/src/main/cpp/skyline/gpu/interconnect/conversion/quads.cpp

@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MPL-2.0
+// Copyright © 2022 Skyline Team and Contributors (https://github.com/skyline-emu/)
+
+#include "quads.h"
+
+namespace skyline::gpu::interconnect::conversion::quads {
+    void GenerateQuadListConversionBuffer(u32 *dest, u32 vertexCount) {
+        #pragma clang loop vectorize(enable) interleave(enable) unroll(enable)
+        for (u32 i{}; i < vertexCount; i += 4) {
+            // Given a quad ABCD, we want to generate triangles ABC & CDA
+            // Triangle ABC
+            *(dest++) = i + 0;
+            *(dest++) = i + 1;
+            *(dest++) = i + 2;
+
+            // Triangle CDA
+            *(dest++) = i + 2;
+            *(dest++) = i + 3;
+            *(dest++) = i + 0;
+        }
+    }
+}

app/src/main/cpp/skyline/gpu/interconnect/conversion/quads.h

@@ -0,0 +1,33 @@
+// SPDX-License-Identifier: MPL-2.0
+// Copyright © 2022 Skyline Team and Contributors (https://github.com/skyline-emu/)
+
+#pragma once
+
+#include <common/base.h>
+
+namespace skyline::gpu::interconnect::conversion::quads {
+    constexpr u32 EmittedIndexCount{6}; //!< The number of indices needed to draw a quad with two triangles
+    constexpr u32 QuadVertexCount{4}; //!< The amount of vertices a quad is composed of
+
+    /**
+     * @return The amount of indices emitted converting a buffer with the supplied element count
+     */
+    constexpr u32 GetIndexCount(u32 count) {
+        return (count * EmittedIndexCount) / QuadVertexCount;
+    }
+
+    /**
+     * @return The minimum size (in bytes) required to store the quad index buffer of the given type after conversion
+     * @tparam T The type of an element in the index buffer
+     */
+    template<typename T>
+    constexpr size_t GetRequiredBufferSize(u32 count) {
+        return GetIndexCount(count) * sizeof(T);
+    }
+
+    /**
+     * @brief Create an index buffer that repeats quad vertices to generate a triangle list
+     * @note The size of the supplied buffer should be at least the size returned by GetRequiredBufferSize()
+     */
+    void GenerateQuadListConversionBuffer(u32 *dest, u32 vertexCount);
+}

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

@@ -15,6 +15,7 @@
 #include "command_executor.h"
 #include "types/tsc.h"
 #include "types/tic.h"
+#include "conversion/quads.h"
 
 namespace skyline::gpu::interconnect {
     namespace maxwell3d = soc::gm20b::engine::maxwell3d::type;
@@ -1581,6 +1582,25 @@ namespace skyline::gpu::interconnect {
         };
         std::array<VertexAttribute, maxwell3d::VertexAttributeCount> vertexAttributes{};
 
+      public:
+        bool needsQuadConversion{}; //!< Whether the current primitive topology is quads and needs conversion to triangles
+
+      private:
+        std::shared_ptr<Buffer> quadListConversionBuffer{}; //!< Index buffer used for QuadList conversion
+
+        /**
+         * @brief Retrieves an index buffer for converting a non-indexed quad list to a triangle list
+         * @result A tuple containing a view over the index buffer, the index type and the index count
+         */
+        std::tuple<BufferView, vk::IndexType, u32> GetQuadListConversionBuffer(u32 count) {
+            vk::DeviceSize size{conversion::quads::GetRequiredBufferSize<u32>(count)};
+            if (!quadListConversionBuffer || quadListConversionBuffer->GetBackingSpan().size_bytes() < size) {
+                quadListConversionBuffer = std::make_shared<Buffer>(gpu, size);
+                conversion::quads::GenerateQuadListConversionBuffer(quadListConversionBuffer->GetBackingSpan().cast<u32>().data(), count);
+            }
+            return {quadListConversionBuffer->GetView(0, size), vk::IndexType::eUint32, conversion::quads::GetIndexCount(count)};
+        }
+
       public:
         void SetVertexBufferStride(u32 index, u32 stride) {
             vertexBuffers[index].bindingDescription.stride = stride;
@@ -1774,27 +1794,29 @@ namespace skyline::gpu::interconnect {
 
       public:
         void SetPrimitiveTopology(maxwell3d::PrimitiveTopology topology) {
-            auto[vkTopology, shaderTopology] = [topology]() -> std::tuple<vk::PrimitiveTopology, ShaderCompiler::InputTopology> {
+            auto[vkTopology, shaderTopology, isQuad] = [topology]() -> std::tuple<vk::PrimitiveTopology, ShaderCompiler::InputTopology, bool> {
                 using MaxwellTopology = maxwell3d::PrimitiveTopology;
                 using VkTopology = vk::PrimitiveTopology;
                 using ShaderTopology = ShaderCompiler::InputTopology;
                 switch (topology) {
                     // @fmt:off
 
-                    case MaxwellTopology::PointList: return {VkTopology::ePointList, ShaderTopology::Points};
+                    case MaxwellTopology::PointList: return {VkTopology::ePointList, ShaderTopology::Points, false};
 
-                    case MaxwellTopology::LineList: return {VkTopology::eLineList, ShaderTopology::Lines};
+                    case MaxwellTopology::LineList: return {VkTopology::eLineList, ShaderTopology::Lines, false};
-                    case MaxwellTopology::LineStrip: return {VkTopology::eLineStrip, ShaderTopology::Lines};
+                    case MaxwellTopology::LineStrip: return {VkTopology::eLineStrip, ShaderTopology::Lines, false};
-                    case MaxwellTopology::LineListWithAdjacency: return {VkTopology::eLineListWithAdjacency, ShaderTopology::LinesAdjacency};
+                    case MaxwellTopology::LineListWithAdjacency: return {VkTopology::eLineListWithAdjacency, ShaderTopology::LinesAdjacency, false};
-                    case MaxwellTopology::LineStripWithAdjacency: return {VkTopology::eLineStripWithAdjacency, ShaderTopology::LinesAdjacency};
+                    case MaxwellTopology::LineStripWithAdjacency: return {VkTopology::eLineStripWithAdjacency, ShaderTopology::LinesAdjacency, false};
 
-                    case MaxwellTopology::TriangleList: return {VkTopology::eTriangleList, ShaderTopology::Triangles};
+                    case MaxwellTopology::TriangleList: return {VkTopology::eTriangleList, ShaderTopology::Triangles, false};
-                    case MaxwellTopology::TriangleStrip: return {VkTopology::eTriangleStrip, ShaderTopology::Triangles};
+                    case MaxwellTopology::TriangleStrip: return {VkTopology::eTriangleStrip, ShaderTopology::Triangles, false};
-                    case MaxwellTopology::TriangleFan: return {VkTopology::eTriangleFan, ShaderTopology::Triangles};
+                    case MaxwellTopology::TriangleFan: return {VkTopology::eTriangleFan, ShaderTopology::Triangles, false};
-                    case MaxwellTopology::TriangleListWithAdjacency: return {VkTopology::eTriangleListWithAdjacency, ShaderTopology::TrianglesAdjacency};
+                    case MaxwellTopology::TriangleListWithAdjacency: return {VkTopology::eTriangleListWithAdjacency, ShaderTopology::TrianglesAdjacency, false};
-                    case MaxwellTopology::TriangleStripWithAdjacency: return {VkTopology::eTriangleStripWithAdjacency, ShaderTopology::TrianglesAdjacency};
+                    case MaxwellTopology::TriangleStripWithAdjacency: return {VkTopology::eTriangleStripWithAdjacency, ShaderTopology::TrianglesAdjacency, false};
 
-                    case MaxwellTopology::PatchList: return {VkTopology::ePatchList, ShaderTopology::Triangles};
+                    case MaxwellTopology::QuadList: return {VkTopology::eTriangleList, ShaderTopology::Triangles, true};
+
+                    case MaxwellTopology::PatchList: return {VkTopology::ePatchList, ShaderTopology::Triangles, false};
 
                     // @fmt:on
 
@@ -1804,6 +1826,7 @@ namespace skyline::gpu::interconnect {
             }();
 
             inputAssemblyState.topology = vkTopology;
+            needsQuadConversion = isQuad;
             UpdateRuntimeInformation(runtimeInfo.input_topology, shaderTopology, maxwell3d::PipelineStage::Geometry);
         }
 
@@ -2595,6 +2618,18 @@ namespace skyline::gpu::interconnect {
             auto boundIndexBuffer{std::make_shared<BoundIndexBuffer>()};
             if constexpr (IsIndexed) {
                 auto indexBufferView{GetIndexBuffer(count)};
+
+                if (needsQuadConversion) {
+                    if (indexBufferView) {
+                        throw exception("Indexed quad conversion is not supported");
+                    } else {
+                        auto[bufferView, indexType, indexCount] = GetQuadListConversionBuffer(count);
+                        indexBufferView = bufferView;
+                        indexBuffer.type = indexType;
+                        count = indexCount;
+                    }
+                }
+
                 {
                     std::scoped_lock lock(indexBufferView);
 

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

@@ -618,6 +618,9 @@ namespace skyline::soc::gm20b::engine::maxwell3d {
             })
 
             ENGINE_CASE(drawVertexCount, {
+                if (context.needsQuadConversion)
+                    context.DrawIndexed(drawVertexCount, *registers.drawVertexFirst, 0);
+                else
                     context.DrawVertex(drawVertexCount, *registers.drawVertexFirst);
             })