Separate GPU engines (part 2/2) (#2440)

* 3D engine now uses DeviceState too, plus new state modification tracking

* Remove old methods code

* Remove GpuState and friends

* Optimize DeviceState, force inline some functions

* This change was not supposed to go in

* Proper channel initialization

* Optimize state read/write methods even more

* Fix debug build

* Do not dirty state if the write is redundant

* The YControl register should dirty either the viewport or front face state too, to update the host origin

* Avoid redundant vertex buffer updates

* Move state and get rid of the Ryujinx.Graphics.Gpu.State namespace

* Comments and nits

* Fix rebase

* PR feedback

* Move changed = false to improve codegen

* PR feedback

* Carry RyuJIT a bit more

Ryujinx.Graphics.Gpu/Engine/Threed/ConditionalRendering.cs

@@ -0,0 +1,130 @@
+using Ryujinx.Common.Logging;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.Types;
+using Ryujinx.Graphics.Gpu.Memory;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Helper methods used for conditional rendering.
+    /// </summary>
+    static class ConditionalRendering
+    {
+        /// <summary>
+        /// Checks if draws and clears should be performed, according
+        /// to currently set conditional rendering conditions.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        /// <param name="memoryManager">Memory manager bound to the channel currently executing</param>
+        /// <param name="address">Conditional rendering buffer address</param>
+        /// <param name="condition">Conditional rendering condition</param>
+        /// <returns>True if rendering is enabled, false otherwise</returns>
+        public static ConditionalRenderEnabled GetRenderEnable(GpuContext context, MemoryManager memoryManager, GpuVa address, Condition condition)
+        {
+            switch (condition)
+            {
+                case Condition.Always:
+                    return ConditionalRenderEnabled.True;
+                case Condition.Never:
+                    return ConditionalRenderEnabled.False;
+                case Condition.ResultNonZero:
+                    return CounterNonZero(context, memoryManager, address.Pack());
+                case Condition.Equal:
+                    return CounterCompare(context, memoryManager, address.Pack(), true);
+                case Condition.NotEqual:
+                    return CounterCompare(context, memoryManager, address.Pack(), false);
+            }
+
+            Logger.Warning?.Print(LogClass.Gpu, $"Invalid conditional render condition \"{condition}\".");
+
+            return ConditionalRenderEnabled.True;
+        }
+
+        /// <summary>
+        /// Checks if the counter value at a given GPU memory address is non-zero.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        /// <param name="memoryManager">Memory manager bound to the channel currently executing</param>
+        /// <param name="gpuVa">GPU virtual address of the counter value</param>
+        /// <returns>True if the value is not zero, false otherwise. Returns host if handling with host conditional rendering</returns>
+        private static ConditionalRenderEnabled CounterNonZero(GpuContext context, MemoryManager memoryManager, ulong gpuVa)
+        {
+            ICounterEvent evt = memoryManager.CounterCache.FindEvent(gpuVa);
+
+            if (evt == null)
+            {
+                return ConditionalRenderEnabled.False;
+            }
+
+            if (context.Renderer.Pipeline.TryHostConditionalRendering(evt, 0L, false))
+            {
+                return ConditionalRenderEnabled.Host;
+            }
+            else
+            {
+                evt.Flush();
+                return (memoryManager.Read<ulong>(gpuVa) != 0) ? ConditionalRenderEnabled.True : ConditionalRenderEnabled.False;
+            }
+        }
+
+        /// <summary>
+        /// Checks if the counter at a given GPU memory address passes a specified equality comparison.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        /// <param name="memoryManager">Memory manager bound to the channel currently executing</param>
+        /// <param name="gpuVa">GPU virtual address</param>
+        /// <param name="isEqual">True to check if the values are equal, false to check if they are not equal</param>
+        /// <returns>True if the condition is met, false otherwise. Returns host if handling with host conditional rendering</returns>
+        private static ConditionalRenderEnabled CounterCompare(GpuContext context, MemoryManager memoryManager, ulong gpuVa, bool isEqual)
+        {
+            ICounterEvent evt = FindEvent(memoryManager.CounterCache, gpuVa);
+            ICounterEvent evt2 = FindEvent(memoryManager.CounterCache, gpuVa + 16);
+
+            bool useHost;
+
+            if (evt != null && evt2 == null)
+            {
+                useHost = context.Renderer.Pipeline.TryHostConditionalRendering(evt, memoryManager.Read<ulong>(gpuVa + 16), isEqual);
+            }
+            else if (evt == null && evt2 != null)
+            {
+                useHost = context.Renderer.Pipeline.TryHostConditionalRendering(evt2, memoryManager.Read<ulong>(gpuVa), isEqual);
+            }
+            else if (evt != null && evt2 != null)
+            {
+                useHost = context.Renderer.Pipeline.TryHostConditionalRendering(evt, evt2, isEqual);
+            }
+            else
+            {
+                useHost = false;
+            }
+
+            if (useHost)
+            {
+                return ConditionalRenderEnabled.Host;
+            }
+            else
+            {
+                evt?.Flush();
+                evt2?.Flush();
+
+                ulong x = memoryManager.Read<ulong>(gpuVa);
+                ulong y = memoryManager.Read<ulong>(gpuVa + 16);
+
+                return (isEqual ? x == y : x != y) ? ConditionalRenderEnabled.True : ConditionalRenderEnabled.False;
+            }
+        }
+
+        /// <summary>
+        /// Tries to find a counter that is supposed to be written at the specified address,
+        /// returning the related event.
+        /// </summary>
+        /// <param name="counterCache">GPU counter cache to search on</param>
+        /// <param name="gpuVa">GPU virtual address where the counter is supposed to be written</param>
+        /// <returns>The counter event, or null if not present</returns>
+        private static ICounterEvent FindEvent(CounterCache counterCache, ulong gpuVa)
+        {
+            return counterCache.FindEvent(gpuVa);
+        }
+    }
+}

Ryujinx.Graphics.Gpu/Engine/Threed/ConstantBufferUpdater.cs

@@ -0,0 +1,173 @@
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Constant buffer updater.
+    /// </summary>
+    class ConstantBufferUpdater
+    {
+        private readonly GpuChannel _channel;
+        private readonly DeviceStateWithShadow<ThreedClassState> _state;
+
+        // State associated with direct uniform buffer updates.
+        // This state is used to attempt to batch together consecutive updates.
+        private ulong _ubBeginCpuAddress = 0;
+        private ulong _ubFollowUpAddress = 0;
+        private ulong _ubByteCount = 0;
+
+        /// <summary>
+        /// Creates a new instance of the constant buffer updater.
+        /// </summary>
+        /// <param name="channel">GPU channel</param>
+        /// <param name="state">Channel state</param>
+        public ConstantBufferUpdater(GpuChannel channel, DeviceStateWithShadow<ThreedClassState> state)
+        {
+            _channel = channel;
+            _state = state;
+        }
+
+        /// <summary>
+        /// Binds a uniform buffer for the vertex shader stage.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void BindVertex(int argument)
+        {
+            Bind(argument, ShaderType.Vertex);
+        }
+
+        /// <summary>
+        /// Binds a uniform buffer for the tessellation control shader stage.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void BindTessControl(int argument)
+        {
+            Bind(argument, ShaderType.TessellationControl);
+        }
+
+        /// <summary>
+        /// Binds a uniform buffer for the tessellation evaluation shader stage.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void BindTessEvaluation(int argument)
+        {
+            Bind(argument, ShaderType.TessellationEvaluation);
+        }
+
+        /// <summary>
+        /// Binds a uniform buffer for the geometry shader stage.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void BindGeometry(int argument)
+        {
+            Bind(argument, ShaderType.Geometry);
+        }
+
+        /// <summary>
+        /// Binds a uniform buffer for the fragment shader stage.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void BindFragment(int argument)
+        {
+            Bind(argument, ShaderType.Fragment);
+        }
+
+        /// <summary>
+        /// Binds a uniform buffer for the specified shader stage.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        /// <param name="type">Shader stage that will access the uniform buffer</param>
+        private void Bind(int argument, ShaderType type)
+        {
+            bool enable = (argument & 1) != 0;
+
+            int index = (argument >> 4) & 0x1f;
+
+            FlushUboDirty();
+
+            if (enable)
+            {
+                var uniformBuffer = _state.State.UniformBufferState;
+
+                ulong address = uniformBuffer.Address.Pack();
+
+                _channel.BufferManager.SetGraphicsUniformBuffer((int)type, index, address, (uint)uniformBuffer.Size);
+            }
+            else
+            {
+                _channel.BufferManager.SetGraphicsUniformBuffer((int)type, index, 0, 0);
+            }
+        }
+
+        /// <summary>
+        /// Flushes any queued UBO updates.
+        /// </summary>
+        public void FlushUboDirty()
+        {
+            if (_ubFollowUpAddress != 0)
+            {
+                var memoryManager = _channel.MemoryManager;
+                memoryManager.Physical.BufferCache.ForceDirty(memoryManager, _ubFollowUpAddress - _ubByteCount, _ubByteCount);
+
+                _ubFollowUpAddress = 0;
+            }
+        }
+
+        /// <summary>
+        /// Updates the uniform buffer data with inline data.
+        /// </summary>
+        /// <param name="argument">New uniform buffer data word</param>
+        public void Update(int argument)
+        {
+            var uniformBuffer = _state.State.UniformBufferState;
+
+            ulong address = uniformBuffer.Address.Pack() + (uint)uniformBuffer.Offset;
+
+            if (_ubFollowUpAddress != address)
+            {
+                FlushUboDirty();
+
+                _ubByteCount = 0;
+                _ubBeginCpuAddress = _channel.MemoryManager.Translate(address);
+            }
+
+            var byteData = MemoryMarshal.Cast<int, byte>(MemoryMarshal.CreateSpan(ref argument, 1));
+            _channel.MemoryManager.Physical.WriteUntracked(_ubBeginCpuAddress + _ubByteCount, byteData);
+
+            _ubFollowUpAddress = address + 4;
+            _ubByteCount += 4;
+
+            _state.State.UniformBufferState.Offset += 4;
+        }
+
+        /// <summary>
+        /// Updates the uniform buffer data with inline data.
+        /// </summary>
+        /// <param name="data">Data to be written to the uniform buffer</param>
+        public void Update(ReadOnlySpan<int> data)
+        {
+            var uniformBuffer = _state.State.UniformBufferState;
+
+            ulong address = uniformBuffer.Address.Pack() + (uint)uniformBuffer.Offset;
+
+            ulong size = (ulong)data.Length * 4;
+
+            if (_ubFollowUpAddress != address)
+            {
+                FlushUboDirty();
+
+                _ubByteCount = 0;
+                _ubBeginCpuAddress = _channel.MemoryManager.Translate(address);
+            }
+
+            var byteData = MemoryMarshal.Cast<int, byte>(data);
+            _channel.MemoryManager.Physical.WriteUntracked(_ubBeginCpuAddress + _ubByteCount, byteData);
+
+            _ubFollowUpAddress = address + size;
+            _ubByteCount += size;
+
+            _state.State.UniformBufferState.Offset += data.Length * 4;
+        }
+    }
+}

Ryujinx.Graphics.Gpu/Engine/Threed/DrawManager.cs

@@ -0,0 +1,410 @@
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.Types;
+using System.Text;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Draw manager.
+    /// </summary>
+    class DrawManager
+    {
+        private readonly GpuContext _context;
+        private readonly GpuChannel _channel;
+        private readonly DeviceStateWithShadow<ThreedClassState> _state;
+        private readonly DrawState _drawState;
+
+        private bool _instancedDrawPending;
+        private bool _instancedIndexed;
+
+        private int _instancedFirstIndex;
+        private int _instancedFirstVertex;
+        private int _instancedFirstInstance;
+        private int _instancedIndexCount;
+        private int _instancedDrawStateFirst;
+        private int _instancedDrawStateCount;
+
+        private int _instanceIndex;
+
+        /// <summary>
+        /// Creates a new instance of the draw manager.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        /// <param name="channel">GPU channel</param>
+        /// <param name="state">Channel state</param>
+        /// <param name="drawState">Draw state</param>
+        public DrawManager(GpuContext context, GpuChannel channel, DeviceStateWithShadow<ThreedClassState> state, DrawState drawState)
+        {
+            _context = context;
+            _channel = channel;
+            _state = state;
+            _drawState = drawState;
+        }
+
+        /// <summary>
+        /// Pushes four 8-bit index buffer elements.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void VbElementU8(int argument)
+        {
+            _drawState.IbStreamer.VbElementU8(_context.Renderer, argument);
+        }
+
+        /// <summary>
+        /// Pushes two 16-bit index buffer elements.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void VbElementU16(int argument)
+        {
+            _drawState.IbStreamer.VbElementU16(_context.Renderer, argument);
+        }
+
+        /// <summary>
+        /// Pushes one 32-bit index buffer element.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void VbElementU32(int argument)
+        {
+            _drawState.IbStreamer.VbElementU32(_context.Renderer, argument);
+        }
+
+        /// <summary>
+        /// Finishes the draw call.
+        /// This draws geometry on the bound buffers based on the current GPU state.
+        /// </summary>
+        /// <param name="engine">3D engine where this method is being called</param>
+        /// <param name="argument">Method call argument</param>
+        public void DrawEnd(ThreedClass engine, int argument)
+        {
+            DrawEnd(engine, _state.State.IndexBufferState.First, (int)_state.State.IndexBufferCount);
+        }
+
+        /// <summary>
+        /// Finishes the draw call.
+        /// This draws geometry on the bound buffers based on the current GPU state.
+        /// </summary>
+        /// <param name="engine">3D engine where this method is being called</param>
+        /// <param name="firstIndex">Index of the first index buffer element used on the draw</param>
+        /// <param name="indexCount">Number of index buffer elements used on the draw</param>
+        private void DrawEnd(ThreedClass engine, int firstIndex, int indexCount)
+        {
+            ConditionalRenderEnabled renderEnable = ConditionalRendering.GetRenderEnable(
+                _context,
+                _channel.MemoryManager,
+                _state.State.RenderEnableAddress,
+                _state.State.RenderEnableCondition);
+
+            if (renderEnable == ConditionalRenderEnabled.False || _instancedDrawPending)
+            {
+                if (renderEnable == ConditionalRenderEnabled.False)
+                {
+                    PerformDeferredDraws();
+                }
+
+                _drawState.DrawIndexed = false;
+
+                if (renderEnable == ConditionalRenderEnabled.Host)
+                {
+                    _context.Renderer.Pipeline.EndHostConditionalRendering();
+                }
+
+                return;
+            }
+
+            _drawState.FirstIndex = firstIndex;
+            _drawState.IndexCount = indexCount;
+
+            engine.UpdateState();
+
+            bool instanced = _drawState.VsUsesInstanceId || _drawState.IsAnyVbInstanced;
+
+            if (instanced)
+            {
+                _instancedDrawPending = true;
+
+                _instancedIndexed = _drawState.DrawIndexed;
+
+                _instancedFirstIndex = firstIndex;
+                _instancedFirstVertex = (int)_state.State.FirstVertex;
+                _instancedFirstInstance = (int)_state.State.FirstInstance;
+
+                _instancedIndexCount = indexCount;
+
+                var drawState = _state.State.VertexBufferDrawState;
+
+                _instancedDrawStateFirst = drawState.First;
+                _instancedDrawStateCount = drawState.Count;
+
+                _drawState.DrawIndexed = false;
+
+                if (renderEnable == ConditionalRenderEnabled.Host)
+                {
+                    _context.Renderer.Pipeline.EndHostConditionalRendering();
+                }
+
+                return;
+            }
+
+            int firstInstance = (int)_state.State.FirstInstance;
+
+            int inlineIndexCount = _drawState.IbStreamer.GetAndResetInlineIndexCount();
+
+            if (inlineIndexCount != 0)
+            {
+                int firstVertex = (int)_state.State.FirstVertex;
+
+                BufferRange br = new BufferRange(_drawState.IbStreamer.GetInlineIndexBuffer(), 0, inlineIndexCount * 4);
+
+                _channel.BufferManager.SetIndexBuffer(br, IndexType.UInt);
+
+                _context.Renderer.Pipeline.DrawIndexed(inlineIndexCount, 1, firstIndex, firstVertex, firstInstance);
+            }
+            else if (_drawState.DrawIndexed)
+            {
+                int firstVertex = (int)_state.State.FirstVertex;
+
+                _context.Renderer.Pipeline.DrawIndexed(indexCount, 1, firstIndex, firstVertex, firstInstance);
+            }
+            else
+            {
+                var drawState = _state.State.VertexBufferDrawState;
+
+                _context.Renderer.Pipeline.Draw(drawState.Count, 1, drawState.First, firstInstance);
+            }
+
+            _drawState.DrawIndexed = false;
+
+            if (renderEnable == ConditionalRenderEnabled.Host)
+            {
+                _context.Renderer.Pipeline.EndHostConditionalRendering();
+            }
+        }
+
+        /// <summary>
+        /// Starts draw.
+        /// This sets primitive type and instanced draw parameters.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void DrawBegin(int argument)
+        {
+            bool incrementInstance = (argument & (1 << 26)) != 0;
+            bool resetInstance = (argument & (1 << 27)) == 0;
+
+            if (_state.State.PrimitiveTypeOverrideEnable)
+            {
+                PrimitiveTypeOverride typeOverride = _state.State.PrimitiveTypeOverride;
+                DrawBegin(incrementInstance, resetInstance, typeOverride.Convert());
+            }
+            else
+            {
+                PrimitiveType type = (PrimitiveType)(argument & 0xffff);
+                DrawBegin(incrementInstance, resetInstance, type.Convert());
+            }
+        }
+
+        /// <summary>
+        /// Starts draw.
+        /// This sets primitive type and instanced draw parameters.
+        /// </summary>
+        /// <param name="incrementInstance">Indicates if the current instance should be incremented</param>
+        /// <param name="resetInstance">Indicates if the current instance should be set to zero</param>
+        /// <param name="topology">Primitive topology</param>
+        private void DrawBegin(bool incrementInstance, bool resetInstance, PrimitiveTopology topology)
+        {
+            if (incrementInstance)
+            {
+                _instanceIndex++;
+            }
+            else if (resetInstance)
+            {
+                PerformDeferredDraws();
+
+                _instanceIndex = 0;
+            }
+
+            _context.Renderer.Pipeline.SetPrimitiveTopology(topology);
+
+            _drawState.Topology = topology;
+        }
+
+        /// <summary>
+        /// Sets the index buffer count.
+        /// This also sets internal state that indicates that the next draw is an indexed draw.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void SetIndexBufferCount(int argument)
+        {
+            _drawState.DrawIndexed = true;
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with a low number of index buffer elements.
+        /// </summary>
+        /// <param name="engine">3D engine where this method is being called</param>
+        /// <param name="argument">Method call argument</param>
+        public void DrawIndexedSmall(ThreedClass engine, int argument)
+        {
+            DrawIndexedSmall(engine, argument, false);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with a low number of index buffer elements.
+        /// </summary>
+        /// <param name="engine">3D engine where this method is being called</param>
+        /// <param name="argument">Method call argument</param>
+        public void DrawIndexedSmall2(ThreedClass engine, int argument)
+        {
+            DrawIndexedSmall(engine, argument);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with a low number of index buffer elements,
+        /// while also pre-incrementing the current instance value.
+        /// </summary>
+        /// <param name="engine">3D engine where this method is being called</param>
+        /// <param name="argument">Method call argument</param>
+        public void DrawIndexedSmallIncInstance(ThreedClass engine, int argument)
+        {
+            DrawIndexedSmall(engine, argument, true);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with a low number of index buffer elements,
+        /// while also pre-incrementing the current instance value.
+        /// </summary>
+        /// <param name="engine">3D engine where this method is being called</param>
+        /// <param name="argument">Method call argument</param>
+        public void DrawIndexedSmallIncInstance2(ThreedClass engine, int argument)
+        {
+            DrawIndexedSmallIncInstance(engine, argument);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with a low number of index buffer elements,
+        /// while optionally also pre-incrementing the current instance value.
+        /// </summary>
+        /// <param name="engine">3D engine where this method is being called</param>
+        /// <param name="argument">Method call argument</param>
+        /// <param name="instanced">True to increment the current instance value, false otherwise</param>
+        private void DrawIndexedSmall(ThreedClass engine, int argument, bool instanced)
+        {
+            PrimitiveTypeOverride typeOverride = _state.State.PrimitiveTypeOverride;
+
+            DrawBegin(instanced, !instanced, typeOverride.Convert());
+
+            int firstIndex = argument & 0xffff;
+            int indexCount = (argument >> 16) & 0xfff;
+
+            bool oldDrawIndexed = _drawState.DrawIndexed;
+
+            _drawState.DrawIndexed = true;
+
+            DrawEnd(engine, firstIndex, indexCount);
+
+            _drawState.DrawIndexed = oldDrawIndexed;
+        }
+
+        /// <summary>
+        /// Perform any deferred draws.
+        /// This is used for instanced draws.
+        /// Since each instance is a separate draw, we defer the draw and accumulate the instance count.
+        /// Once we detect the last instanced draw, then we perform the host instanced draw,
+        /// with the accumulated instance count.
+        /// </summary>
+        public void PerformDeferredDraws()
+        {
+            // Perform any pending instanced draw.
+            if (_instancedDrawPending)
+            {
+                _instancedDrawPending = false;
+
+                if (_instancedIndexed)
+                {
+                    _context.Renderer.Pipeline.DrawIndexed(
+                        _instancedIndexCount,
+                        _instanceIndex + 1,
+                        _instancedFirstIndex,
+                        _instancedFirstVertex,
+                        _instancedFirstInstance);
+                }
+                else
+                {
+                    _context.Renderer.Pipeline.Draw(
+                        _instancedDrawStateCount,
+                        _instanceIndex + 1,
+                        _instancedDrawStateFirst,
+                        _instancedFirstInstance);
+                }
+            }
+        }
+
+        /// <summary>
+        /// Clears the current color and depth-stencil buffers.
+        /// Which buffers should be cleared is also specified on the argument.
+        /// </summary>
+        /// <param name="engine">3D engine where this method is being called</param>
+        /// <param name="argument">Method call argument</param>
+        public void Clear(ThreedClass engine, int argument)
+        {
+            ConditionalRenderEnabled renderEnable = ConditionalRendering.GetRenderEnable(
+                _context,
+                _channel.MemoryManager,
+                _state.State.RenderEnableAddress,
+                _state.State.RenderEnableCondition);
+
+            if (renderEnable == ConditionalRenderEnabled.False)
+            {
+                return;
+            }
+
+            // Scissor and rasterizer discard also affect clears.
+            engine.UpdateState((1UL << StateUpdater.RasterizerStateIndex) | (1UL << StateUpdater.ScissorStateIndex));
+
+            int index = (argument >> 6) & 0xf;
+
+            engine.UpdateRenderTargetState(useControl: false, singleUse: index);
+
+            _channel.TextureManager.UpdateRenderTargets();
+
+            bool clearDepth = (argument & 1) != 0;
+            bool clearStencil = (argument & 2) != 0;
+
+            uint componentMask = (uint)((argument >> 2) & 0xf);
+
+            if (componentMask != 0)
+            {
+                var clearColor = _state.State.ClearColors;
+
+                ColorF color = new ColorF(clearColor.Red, clearColor.Green, clearColor.Blue, clearColor.Alpha);
+
+                _context.Renderer.Pipeline.ClearRenderTargetColor(index, componentMask, color);
+            }
+
+            if (clearDepth || clearStencil)
+            {
+                float depthValue = _state.State.ClearDepthValue;
+                int stencilValue = (int)_state.State.ClearStencilValue;
+
+                int stencilMask = 0;
+
+                if (clearStencil)
+                {
+                    stencilMask = _state.State.StencilTestState.FrontMask;
+                }
+
+                _context.Renderer.Pipeline.ClearRenderTargetDepthStencil(
+                    depthValue,
+                    clearDepth,
+                    stencilValue,
+                    stencilMask);
+            }
+
+            engine.UpdateRenderTargetState(useControl: true);
+
+            if (renderEnable == ConditionalRenderEnabled.Host)
+            {
+                _context.Renderer.Pipeline.EndHostConditionalRendering();
+            }
+        }
+    }
+}

Ryujinx.Graphics.Gpu/Engine/Threed/DrawState.cs

@@ -0,0 +1,45 @@
+using Ryujinx.Graphics.GAL;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Draw state.
+    /// </summary>
+    class DrawState
+    {
+        /// <summary>
+        /// First index to be used for the draw on the index buffer.
+        /// </summary>
+        public int FirstIndex;
+
+        /// <summary>
+        /// Number of indices to be used for the draw on the index buffer.
+        /// </summary>
+        public int IndexCount;
+
+        /// <summary>
+        /// Indicates if the next draw will be a indexed draw.
+        /// </summary>
+        public bool DrawIndexed;
+
+        /// <summary>
+        /// Indicates if any of the currently used vertex shaders reads the instance ID.
+        /// </summary>
+        public bool VsUsesInstanceId;
+
+        /// <summary>
+        /// Indicates if any of the currently used vertex buffers is instanced.
+        /// </summary>
+        public bool IsAnyVbInstanced;
+
+        /// <summary>
+        /// Primitive topology for the next draw.
+        /// </summary>
+        public PrimitiveTopology Topology;
+
+        /// <summary>
+        /// Index buffer data streamer for inline index buffer updates, such as those used in legacy OpenGL.
+        /// </summary>
+        public IbStreamer IbStreamer = new IbStreamer();
+    }
+}

Ryujinx.Graphics.Gpu/Engine/Threed/IbStreamer.cs

@@ -0,0 +1,142 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.GAL;
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Holds inline index buffer state.
+    /// The inline index buffer data is sent to the GPU through the command buffer.
+    /// </summary>
+    struct IbStreamer
+    {
+        private BufferHandle _inlineIndexBuffer;
+        private int _inlineIndexBufferSize;
+        private int _inlineIndexCount;
+
+        /// <summary>
+        /// Indicates if any index buffer data has been pushed.
+        /// </summary>
+        public bool HasInlineIndexData => _inlineIndexCount != 0;
+
+        /// <summary>
+        /// Gets the handle for the host buffer currently holding the inline index buffer data.
+        /// </summary>
+        /// <returns>Host buffer handle</returns>
+        public BufferHandle GetInlineIndexBuffer()
+        {
+            return _inlineIndexBuffer;
+        }
+
+        /// <summary>
+        /// Gets the number of elements on the current inline index buffer,
+        /// while also reseting it to zero for the next draw.
+        /// </summary>
+        /// <returns>Inline index bufffer count</returns>
+        public int GetAndResetInlineIndexCount()
+        {
+            int temp = _inlineIndexCount;
+            _inlineIndexCount = 0;
+            return temp;
+        }
+
+        /// <summary>
+        /// Pushes four 8-bit index buffer elements.
+        /// </summary>
+        /// <param name="renderer">Host renderer</param>
+        /// <param name="argument">Method call argument</param>
+        public void VbElementU8(IRenderer renderer, int argument)
+        {
+            byte i0 = (byte)argument;
+            byte i1 = (byte)(argument >> 8);
+            byte i2 = (byte)(argument >> 16);
+            byte i3 = (byte)(argument >> 24);
+
+            Span<uint> data = stackalloc uint[4];
+
+            data[0] = i0;
+            data[1] = i1;
+            data[2] = i2;
+            data[3] = i3;
+
+            int offset = _inlineIndexCount * 4;
+
+            renderer.SetBufferData(GetInlineIndexBuffer(renderer, offset), offset, MemoryMarshal.Cast<uint, byte>(data));
+
+            _inlineIndexCount += 4;
+        }
+
+        /// <summary>
+        /// Pushes two 16-bit index buffer elements.
+        /// </summary>
+        /// <param name="renderer">Host renderer</param>
+        /// <param name="argument">Method call argument</param>
+        public void VbElementU16(IRenderer renderer, int argument)
+        {
+            ushort i0 = (ushort)argument;
+            ushort i1 = (ushort)(argument >> 16);
+
+            Span<uint> data = stackalloc uint[2];
+
+            data[0] = i0;
+            data[1] = i1;
+
+            int offset = _inlineIndexCount * 4;
+
+            renderer.SetBufferData(GetInlineIndexBuffer(renderer, offset), offset, MemoryMarshal.Cast<uint, byte>(data));
+
+            _inlineIndexCount += 2;
+        }
+
+        /// <summary>
+        /// Pushes one 32-bit index buffer element.
+        /// </summary>
+        /// <param name="renderer">Host renderer</param>
+        /// <param name="argument">Method call argument</param>
+        public void VbElementU32(IRenderer renderer, int argument)
+        {
+            uint i0 = (uint)argument;
+
+            Span<uint> data = stackalloc uint[1];
+
+            data[0] = i0;
+
+            int offset = _inlineIndexCount++ * 4;
+
+            renderer.SetBufferData(GetInlineIndexBuffer(renderer, offset), offset, MemoryMarshal.Cast<uint, byte>(data));
+        }
+
+        /// <summary>
+        /// Gets the handle of a buffer large enough to hold the data that will be written to <paramref name="offset"/>.
+        /// </summary>
+        /// <param name="renderer">Host renderer</param>
+        /// <param name="offset">Offset where the data will be written</param>
+        /// <returns>Buffer handle</returns>
+        private BufferHandle GetInlineIndexBuffer(IRenderer renderer, int offset)
+        {
+            // Calculate a reasonable size for the buffer that can fit all the data,
+            // and that also won't require frequent resizes if we need to push more data.
+            int size = BitUtils.AlignUp(offset + 0x10, 0x200);
+
+            if (_inlineIndexBuffer == BufferHandle.Null)
+            {
+                _inlineIndexBuffer = renderer.CreateBuffer(size);
+                _inlineIndexBufferSize = size;
+            }
+            else if (_inlineIndexBufferSize < size)
+            {
+                BufferHandle oldBuffer = _inlineIndexBuffer;
+                int oldSize = _inlineIndexBufferSize;
+
+                _inlineIndexBuffer = renderer.CreateBuffer(size);
+                _inlineIndexBufferSize = size;
+
+                renderer.Pipeline.CopyBuffer(oldBuffer, _inlineIndexBuffer, 0, 0, oldSize);
+                renderer.DeleteBuffer(oldBuffer);
+            }
+
+            return _inlineIndexBuffer;
+        }
+    }
+}

Ryujinx.Graphics.Gpu/Engine/Threed/SemaphoreUpdater.cs

@@ -0,0 +1,222 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.GAL;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Semaphore updater.
+    /// </summary>
+    class SemaphoreUpdater
+    {
+        private const int NsToTicksFractionNumerator = 384;
+        private const int NsToTicksFractionDenominator = 625;
+
+        /// <summary>
+        /// GPU semaphore operation.
+        /// </summary>
+        private enum SemaphoreOperation
+        {
+            Release = 0,
+            Acquire = 1,
+            Counter = 2
+        }
+
+        /// <summary>
+        /// Counter type for GPU counter reset.
+        /// </summary>
+        private enum ResetCounterType
+        {
+            SamplesPassed = 1,
+            ZcullStats = 2,
+            TransformFeedbackPrimitivesWritten = 0x10,
+            InputVertices = 0x12,
+            InputPrimitives = 0x13,
+            VertexShaderInvocations = 0x15,
+            TessControlShaderInvocations = 0x16,
+            TessEvaluationShaderInvocations = 0x17,
+            TessEvaluationShaderPrimitives = 0x18,
+            GeometryShaderInvocations = 0x1a,
+            GeometryShaderPrimitives = 0x1b,
+            ClipperInputPrimitives = 0x1c,
+            ClipperOutputPrimitives = 0x1d,
+            FragmentShaderInvocations = 0x1e,
+            PrimitivesGenerated = 0x1f
+        }
+
+        /// <summary>
+        /// Counter type for GPU counter reporting.
+        /// </summary>
+        private enum ReportCounterType
+        {
+            Zero = 0,
+            InputVertices = 1,
+            InputPrimitives = 3,
+            VertexShaderInvocations = 5,
+            GeometryShaderInvocations = 7,
+            GeometryShaderPrimitives = 9,
+            ZcullStats0 = 0xa,
+            TransformFeedbackPrimitivesWritten = 0xb,
+            ZcullStats1 = 0xc,
+            ZcullStats2 = 0xe,
+            ClipperInputPrimitives = 0xf,
+            ZcullStats3 = 0x10,
+            ClipperOutputPrimitives = 0x11,
+            PrimitivesGenerated = 0x12,
+            FragmentShaderInvocations = 0x13,
+            SamplesPassed = 0x15,
+            TransformFeedbackOffset = 0x1a,
+            TessControlShaderInvocations = 0x1b,
+            TessEvaluationShaderInvocations = 0x1d,
+            TessEvaluationShaderPrimitives = 0x1f
+        }
+
+        private readonly GpuContext _context;
+        private readonly GpuChannel _channel;
+        private readonly DeviceStateWithShadow<ThreedClassState> _state;
+
+        /// <summary>
+        /// Creates a new instance of the semaphore updater.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        /// <param name="channel">GPU channel</param>
+        /// <param name="state">Channel state</param>
+        public SemaphoreUpdater(GpuContext context, GpuChannel channel, DeviceStateWithShadow<ThreedClassState> state)
+        {
+            _context = context;
+            _channel = channel;
+            _state = state;
+        }
+
+        /// <summary>
+        /// Resets the value of an internal GPU counter back to zero.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void ResetCounter(int argument)
+        {
+            ResetCounterType type = (ResetCounterType)argument;
+
+            switch (type)
+            {
+                case ResetCounterType.SamplesPassed:
+                    _context.Renderer.ResetCounter(CounterType.SamplesPassed);
+                    break;
+                case ResetCounterType.PrimitivesGenerated:
+                    _context.Renderer.ResetCounter(CounterType.PrimitivesGenerated);
+                    break;
+                case ResetCounterType.TransformFeedbackPrimitivesWritten:
+                    _context.Renderer.ResetCounter(CounterType.TransformFeedbackPrimitivesWritten);
+                    break;
+            }
+        }
+
+        /// <summary>
+        /// Writes a GPU counter to guest memory.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void Report(int argument)
+        {
+            SemaphoreOperation op = (SemaphoreOperation)(argument & 3);
+            ReportCounterType type = (ReportCounterType)((argument >> 23) & 0x1f);
+
+            switch (op)
+            {
+                case SemaphoreOperation.Release: ReleaseSemaphore(); break;
+                case SemaphoreOperation.Counter: ReportCounter(type); break;
+            }
+        }
+
+        /// <summary>
+        /// Writes (or Releases) a GPU semaphore value to guest memory.
+        /// </summary>
+        private void ReleaseSemaphore()
+        {
+            _channel.MemoryManager.Write(_state.State.SemaphoreAddress.Pack(), _state.State.SemaphorePayload);
+
+            _context.AdvanceSequence();
+        }
+
+        /// <summary>
+        /// Packed GPU counter data (including GPU timestamp) in memory.
+        /// </summary>
+        private struct CounterData
+        {
+            public ulong Counter;
+            public ulong Timestamp;
+        }
+
+        /// <summary>
+        /// Writes a GPU counter to guest memory.
+        /// This also writes the current timestamp value.
+        /// </summary>
+        /// <param name="type">Counter to be written to memory</param>
+        private void ReportCounter(ReportCounterType type)
+        {
+            ulong gpuVa = _state.State.SemaphoreAddress.Pack();
+
+            ulong ticks = ConvertNanosecondsToTicks((ulong)PerformanceCounter.ElapsedNanoseconds);
+
+            if (GraphicsConfig.FastGpuTime)
+            {
+                // Divide by some amount to report time as if operations were performed faster than they really are.
+                // This can prevent some games from switching to a lower resolution because rendering is too slow.
+                ticks /= 256;
+            }
+
+            ICounterEvent counter = null;
+
+            void resultHandler(object evt, ulong result)
+            {
+                CounterData counterData = new CounterData
+                {
+                    Counter = result,
+                    Timestamp = ticks
+                };
+
+                if (counter?.Invalid != true)
+                {
+                    _channel.MemoryManager.Write(gpuVa, counterData);
+                }
+            }
+
+            switch (type)
+            {
+                case ReportCounterType.Zero:
+                    resultHandler(null, 0);
+                    break;
+                case ReportCounterType.SamplesPassed:
+                    counter = _context.Renderer.ReportCounter(CounterType.SamplesPassed, resultHandler);
+                    break;
+                case ReportCounterType.PrimitivesGenerated:
+                    counter = _context.Renderer.ReportCounter(CounterType.PrimitivesGenerated, resultHandler);
+                    break;
+                case ReportCounterType.TransformFeedbackPrimitivesWritten:
+                    counter = _context.Renderer.ReportCounter(CounterType.TransformFeedbackPrimitivesWritten, resultHandler);
+                    break;
+            }
+
+            _channel.MemoryManager.CounterCache.AddOrUpdate(gpuVa, counter);
+        }
+
+        /// <summary>
+        /// Converts a nanoseconds timestamp value to Maxwell time ticks.
+        /// </summary>
+        /// <remarks>
+        /// The frequency is 614400000 Hz.
+        /// </remarks>
+        /// <param name="nanoseconds">Timestamp in nanoseconds</param>
+        /// <returns>Maxwell ticks</returns>
+        private static ulong ConvertNanosecondsToTicks(ulong nanoseconds)
+        {
+            // We need to divide first to avoid overflows.
+            // We fix up the result later by calculating the difference and adding
+            // that to the result.
+            ulong divided = nanoseconds / NsToTicksFractionDenominator;
+
+            ulong rounded = divided * NsToTicksFractionDenominator;
+
+            ulong errorBias = (nanoseconds - rounded) * NsToTicksFractionNumerator / NsToTicksFractionDenominator;
+
+            return divided * NsToTicksFractionNumerator + errorBias;
+        }
+    }
+}

Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdateTracker.cs

@@ -0,0 +1,166 @@
+using Ryujinx.Graphics.Device;
+using System;
+using System.Collections.Generic;
+using System.Diagnostics;
+using System.Numerics;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// State update callback entry, with the callback function and associated field names.
+    /// </summary>
+    struct StateUpdateCallbackEntry
+    {
+        /// <summary>
+        /// Callback function, to be called if the register was written as the state needs to be updated.
+        /// </summary>
+        public Action Callback { get; }
+
+        /// <summary>
+        /// Name of the state fields (registers) associated with the callback function.
+        /// </summary>
+        public string[] FieldNames { get; }
+
+        /// <summary>
+        /// Creates a new state update callback entry.
+        /// </summary>
+        /// <param name="callback">Callback function, to be called if the register was written as the state needs to be updated</param>
+        /// <param name="fieldNames">Name of the state fields (registers) associated with the callback function</param>
+        public StateUpdateCallbackEntry(Action callback, params string[] fieldNames)
+        {
+            Callback = callback;
+            FieldNames = fieldNames;
+        }
+    }
+
+    /// <summary>
+    /// GPU state update tracker.
+    /// </summary>
+    /// <typeparam name="TState">State type</typeparam>
+    class StateUpdateTracker<TState>
+    {
+        private const int BlockSize = 0xe00;
+        private const int RegisterSize = sizeof(uint);
+
+        private readonly byte[] _registerToGroupMapping;
+        private readonly Action[] _callbacks;
+        private ulong _dirtyMask;
+
+        /// <summary>
+        /// Creates a new instance of the state update tracker.
+        /// </summary>
+        /// <param name="entries">Update tracker callback entries</param>
+        public StateUpdateTracker(StateUpdateCallbackEntry[] entries)
+        {
+            _registerToGroupMapping = new byte[BlockSize];
+            _callbacks = new Action[entries.Length];
+
+            var fieldToDelegate = new Dictionary<string, int>();
+
+            for (int entryIndex = 0; entryIndex < entries.Length; entryIndex++)
+            {
+                var entry = entries[entryIndex];
+
+                foreach (var fieldName in entry.FieldNames)
+                {
+                    fieldToDelegate.Add(fieldName, entryIndex);
+                }
+
+                _callbacks[entryIndex] = entry.Callback;
+            }
+
+            var fields = typeof(TState).GetFields();
+            int offset = 0;
+
+            for (int fieldIndex = 0; fieldIndex < fields.Length; fieldIndex++)
+            {
+                var field = fields[fieldIndex];
+
+                int sizeOfField = SizeCalculator.SizeOf(field.FieldType);
+
+                if (fieldToDelegate.TryGetValue(field.Name, out int entryIndex))
+                {
+                    for (int i = 0; i < ((sizeOfField + 3) & ~3); i += 4)
+                    {
+                        _registerToGroupMapping[(offset + i) / RegisterSize] = (byte)(entryIndex + 1);
+                    }
+                }
+
+                offset += sizeOfField;
+            }
+
+            Debug.Assert(offset == Unsafe.SizeOf<TState>());
+        }
+
+        /// <summary>
+        /// Sets a register as modified.
+        /// </summary>
+        /// <param name="offset">Register offset in bytes</param>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void SetDirty(int offset)
+        {
+            uint index = (uint)offset / RegisterSize;
+
+            if (index < BlockSize)
+            {
+                int groupIndex = Unsafe.Add(ref MemoryMarshal.GetArrayDataReference(_registerToGroupMapping), (IntPtr)index);
+                if (groupIndex != 0)
+                {
+                    groupIndex--;
+                    _dirtyMask |= 1UL << groupIndex;
+                }
+            }
+        }
+
+        /// <summary>
+        /// Forces a register group as dirty, by index.
+        /// </summary>
+        /// <param name="groupIndex">Index of the group to be dirtied</param>
+        public void ForceDirty(int groupIndex)
+        {
+            if ((uint)groupIndex >= _callbacks.Length)
+            {
+                throw new ArgumentOutOfRangeException(nameof(groupIndex));
+            }
+
+            _dirtyMask |= 1UL << groupIndex;
+        }
+
+        /// <summary>
+        /// Forces all register groups as dirty, triggering a full update on the next call to <see cref="Update"/>.
+        /// </summary>
+        public void SetAllDirty()
+        {
+            Debug.Assert(_callbacks.Length <= sizeof(ulong) * 8);
+            _dirtyMask = ulong.MaxValue >> ((sizeof(ulong) * 8) - _callbacks.Length);
+        }
+
+        /// <summary>
+        /// Check all the groups specified by <paramref name="checkMask"/> for modification, and update if modified.
+        /// </summary>
+        /// <param name="checkMask">Mask, where each bit set corresponds to a group index that should be checked</param>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void Update(ulong checkMask)
+        {
+            ulong mask = _dirtyMask & checkMask;
+            if (mask == 0)
+            {
+                return;
+            }
+
+            do
+            {
+                int groupIndex = BitOperations.TrailingZeroCount(mask);
+
+                _callbacks[groupIndex]();
+
+                mask &= ~(1UL << groupIndex);
+            }
+            while (mask != 0);
+
+            _dirtyMask &= ~checkMask;
+        }
+    }
+}

Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClass.cs

@@ -0,0 +1,428 @@
+using Ryujinx.Graphics.Device;
+using Ryujinx.Graphics.Gpu.Engine.InlineToMemory;
+using System;
+using System.Collections.Generic;
+using System.Runtime.CompilerServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Represents a 3D engine class.
+    /// </summary>
+    class ThreedClass : IDeviceState
+    {
+        private readonly GpuContext _context;
+        private readonly DeviceStateWithShadow<ThreedClassState> _state;
+
+        private readonly InlineToMemoryClass _i2mClass;
+        private readonly DrawManager _drawManager;
+        private readonly SemaphoreUpdater _semaphoreUpdater;
+        private readonly ConstantBufferUpdater _cbUpdater;
+        private readonly StateUpdater _stateUpdater;
+
+        /// <summary>
+        /// Creates a new instance of the 3D engine class.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        /// <param name="channel">GPU channel</param>
+        public ThreedClass(GpuContext context, GpuChannel channel)
+        {
+            _context = context;
+            _state = new DeviceStateWithShadow<ThreedClassState>(new Dictionary<string, RwCallback>
+            {
+                { nameof(ThreedClassState.LaunchDma), new RwCallback(LaunchDma, null) },
+                { nameof(ThreedClassState.LoadInlineData), new RwCallback(LoadInlineData, null) },
+                { nameof(ThreedClassState.SyncpointAction), new RwCallback(IncrementSyncpoint, null) },
+                { nameof(ThreedClassState.TextureBarrier), new RwCallback(TextureBarrier, null) },
+                { nameof(ThreedClassState.TextureBarrierTiled), new RwCallback(TextureBarrierTiled, null) },
+                { nameof(ThreedClassState.VbElementU8), new RwCallback(VbElementU8, null) },
+                { nameof(ThreedClassState.VbElementU16), new RwCallback(VbElementU16, null) },
+                { nameof(ThreedClassState.VbElementU32), new RwCallback(VbElementU32, null) },
+                { nameof(ThreedClassState.ResetCounter), new RwCallback(ResetCounter, null) },
+                { nameof(ThreedClassState.RenderEnableCondition), new RwCallback(null, Zero) },
+                { nameof(ThreedClassState.DrawEnd), new RwCallback(DrawEnd, null) },
+                { nameof(ThreedClassState.DrawBegin), new RwCallback(DrawBegin, null) },
+                { nameof(ThreedClassState.DrawIndexedSmall), new RwCallback(DrawIndexedSmall, null) },
+                { nameof(ThreedClassState.DrawIndexedSmall2), new RwCallback(DrawIndexedSmall2, null) },
+                { nameof(ThreedClassState.DrawIndexedSmallIncInstance), new RwCallback(DrawIndexedSmallIncInstance, null) },
+                { nameof(ThreedClassState.DrawIndexedSmallIncInstance2), new RwCallback(DrawIndexedSmallIncInstance2, null) },
+                { nameof(ThreedClassState.IndexBufferCount), new RwCallback(SetIndexBufferCount, null) },
+                { nameof(ThreedClassState.Clear), new RwCallback(Clear, null) },
+                { nameof(ThreedClassState.SemaphoreControl), new RwCallback(Report, null) },
+                { nameof(ThreedClassState.SetFalcon04), new RwCallback(SetFalcon04, null) },
+                { nameof(ThreedClassState.UniformBufferUpdateData), new RwCallback(ConstantBufferUpdate, null) },
+                { nameof(ThreedClassState.UniformBufferBindVertex), new RwCallback(ConstantBufferBindVertex, null) },
+                { nameof(ThreedClassState.UniformBufferBindTessControl), new RwCallback(ConstantBufferBindTessControl, null) },
+                { nameof(ThreedClassState.UniformBufferBindTessEvaluation), new RwCallback(ConstantBufferBindTessEvaluation, null) },
+                { nameof(ThreedClassState.UniformBufferBindGeometry), new RwCallback(ConstantBufferBindGeometry, null) },
+                { nameof(ThreedClassState.UniformBufferBindFragment), new RwCallback(ConstantBufferBindFragment, null) }
+            });
+
+            _i2mClass = new InlineToMemoryClass(context, channel, initializeState: false);
+
+            var drawState = new DrawState();
+
+            _drawManager = new DrawManager(context, channel, _state, drawState);
+            _semaphoreUpdater = new SemaphoreUpdater(context, channel, _state);
+            _cbUpdater = new ConstantBufferUpdater(channel, _state);
+            _stateUpdater = new StateUpdater(context, channel, _state, drawState);
+
+            // This defaults to "always", even without any register write.
+            // Reads just return 0, regardless of what was set there.
+            _state.State.RenderEnableCondition = Condition.Always;
+        }
+
+        /// <summary>
+        /// Reads data from the class registers.
+        /// </summary>
+        /// <param name="offset">Register byte offset</param>
+        /// <returns>Data at the specified offset</returns>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public int Read(int offset) => _state.Read(offset);
+
+        /// <summary>
+        /// Writes data to the class registers.
+        /// </summary>
+        /// <param name="offset">Register byte offset</param>
+        /// <param name="data">Data to be written</param>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void Write(int offset, int data)
+        {
+            _state.WriteWithRedundancyCheck(offset, data, out bool valueChanged);
+
+            if (valueChanged)
+            {
+                _stateUpdater.SetDirty(offset);
+            }
+        }
+
+        /// <summary>
+        /// Sets the shadow ram control value of all sub-channels.
+        /// </summary>
+        /// <param name="control">New shadow ram control value</param>
+        public void SetShadowRamControl(int control)
+        {
+            _state.State.SetMmeShadowRamControl = (uint)control;
+        }
+
+        /// <summary>
+        /// Updates current host state for all registers modified since the last call to this method.
+        /// </summary>
+        public void UpdateState()
+        {
+            _cbUpdater.FlushUboDirty();
+            _stateUpdater.Update();
+        }
+
+        /// <summary>
+        /// Updates current host state for all registers modified since the last call to this method.
+        /// </summary>
+        /// <param name="mask">Mask where each bit set indicates that the respective state group index should be checked</param>
+        public void UpdateState(ulong mask)
+        {
+            _stateUpdater.Update(mask);
+        }
+
+        /// <summary>
+        /// Updates render targets (color and depth-stencil buffers) based on current render target state.
+        /// </summary>
+        /// <param name="useControl">Use draw buffers information from render target control register</param>
+        /// <param name="singleUse">If this is not -1, it indicates that only the given indexed target will be used.</param>
+        public void UpdateRenderTargetState(bool useControl, int singleUse = -1)
+        {
+            _stateUpdater.UpdateRenderTargetState(useControl, singleUse);
+        }
+
+        /// <summary>
+        /// Marks the entire state as dirty, forcing a full host state update before the next draw.
+        /// </summary>
+        public void ForceStateDirty()
+        {
+            _stateUpdater.SetAllDirty();
+        }
+
+        /// <summary>
+        /// Forces the shaders to be rebound on the next draw.
+        /// </summary>
+        public void ForceShaderUpdate()
+        {
+            _stateUpdater.ForceShaderUpdate();
+        }
+
+        /// <summary>
+        /// Flushes any queued UBO updates.
+        /// </summary>
+        public void FlushUboDirty()
+        {
+            _cbUpdater.FlushUboDirty();
+        }
+
+        /// <summary>
+        /// Perform any deferred draws.
+        /// </summary>
+        public void PerformDeferredDraws()
+        {
+            _drawManager.PerformDeferredDraws();
+        }
+
+        /// <summary>
+        /// Updates the currently bound constant buffer.
+        /// </summary>
+        /// <param name="data">Data to be written to the buffer</param>
+        public void ConstantBufferUpdate(ReadOnlySpan<int> data)
+        {
+            _cbUpdater.Update(data);
+        }
+
+        /// <summary>
+        /// Launches the Inline-to-Memory DMA copy operation.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void LaunchDma(int argument)
+        {
+            _i2mClass.LaunchDma(ref Unsafe.As<ThreedClassState, InlineToMemoryClassState>(ref _state.State), argument);
+        }
+
+        /// <summary>
+        /// Pushes a word of data to the Inline-to-Memory engine.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void LoadInlineData(int argument)
+        {
+            _i2mClass.LoadInlineData(argument);
+        }
+
+        /// <summary>
+        /// Performs an incrementation on a syncpoint.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void IncrementSyncpoint(int argument)
+        {
+            uint syncpointId = (uint)argument & 0xFFFF;
+
+            _context.CreateHostSyncIfNeeded();
+            _context.Renderer.UpdateCounters(); // Poll the query counters, the game may want an updated result.
+            _context.Synchronization.IncrementSyncpoint(syncpointId);
+        }
+
+        /// <summary>
+        /// Issues a texture barrier.
+        /// This waits until previous texture writes from the GPU to finish, before
+        /// performing new operations with said textures.
+        /// </summary>
+        /// <param name="argument">Method call argument (unused)</param>
+        private void TextureBarrier(int argument)
+        {
+            _context.Renderer.Pipeline.TextureBarrier();
+        }
+
+        /// <summary>
+        /// Issues a texture barrier.
+        /// This waits until previous texture writes from the GPU to finish, before
+        /// performing new operations with said textures.
+        /// This performs a per-tile wait, it is only valid if both the previous write
+        /// and current access has the same access patterns.
+        /// This may be faster than the regular barrier on tile-based rasterizers.
+        /// </summary>
+        /// <param name="argument">Method call argument (unused)</param>
+        private void TextureBarrierTiled(int argument)
+        {
+            _context.Renderer.Pipeline.TextureBarrierTiled();
+        }
+
+        /// <summary>
+        /// Pushes four 8-bit index buffer elements.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void VbElementU8(int argument)
+        {
+            _drawManager.VbElementU8(argument);
+        }
+
+        /// <summary>
+        /// Pushes two 16-bit index buffer elements.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void VbElementU16(int argument)
+        {
+            _drawManager.VbElementU16(argument);
+        }
+
+        /// <summary>
+        /// Pushes one 32-bit index buffer element.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void VbElementU32(int argument)
+        {
+            _drawManager.VbElementU32(argument);
+        }
+
+        /// <summary>
+        /// Resets the value of an internal GPU counter back to zero.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void ResetCounter(int argument)
+        {
+            _semaphoreUpdater.ResetCounter(argument);
+        }
+
+        /// <summary>
+        /// Finishes the draw call.
+        /// This draws geometry on the bound buffers based on the current GPU state.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawEnd(int argument)
+        {
+            _drawManager.DrawEnd(this, argument);
+        }
+
+        /// <summary>
+        /// Starts draw.
+        /// This sets primitive type and instanced draw parameters.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawBegin(int argument)
+        {
+            _drawManager.DrawBegin(argument);
+        }
+
+        /// <summary>
+        /// Sets the index buffer count.
+        /// This also sets internal state that indicates that the next draw is an indexed draw.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void SetIndexBufferCount(int argument)
+        {
+            _drawManager.SetIndexBufferCount(argument);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with a low number of index buffer elements.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawIndexedSmall(int argument)
+        {
+            _drawManager.DrawIndexedSmall(this, argument);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with a low number of index buffer elements.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawIndexedSmall2(int argument)
+        {
+            _drawManager.DrawIndexedSmall2(this, argument);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with a low number of index buffer elements,
+        /// while also pre-incrementing the current instance value.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawIndexedSmallIncInstance(int argument)
+        {
+            _drawManager.DrawIndexedSmallIncInstance(this, argument);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with a low number of index buffer elements,
+        /// while also pre-incrementing the current instance value.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawIndexedSmallIncInstance2(int argument)
+        {
+            _drawManager.DrawIndexedSmallIncInstance2(this, argument);
+        }
+
+        /// <summary>
+        /// Clears the current color and depth-stencil buffers.
+        /// Which buffers should be cleared is also specified on the argument.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void Clear(int argument)
+        {
+            _drawManager.Clear(this, argument);
+        }
+
+        /// <summary>
+        /// Writes a GPU counter to guest memory.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void Report(int argument)
+        {
+            _semaphoreUpdater.Report(argument);
+        }
+
+        /// <summary>
+        /// Performs high-level emulation of Falcon microcode function number "4".
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void SetFalcon04(int argument)
+        {
+            _state.State.SetMmeShadowScratch[0] = 1;
+        }
+
+        /// <summary>
+        /// Updates the uniform buffer data with inline data.
+        /// </summary>
+        /// <param name="argument">New uniform buffer data word</param>
+        private void ConstantBufferUpdate(int argument)
+        {
+            _cbUpdater.Update(argument);
+        }
+
+        /// <summary>
+        /// Binds a uniform buffer for the vertex shader stage.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void ConstantBufferBindVertex(int argument)
+        {
+            _cbUpdater.BindVertex(argument);
+        }
+
+        /// <summary>
+        /// Binds a uniform buffer for the tessellation control shader stage.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void ConstantBufferBindTessControl(int argument)
+        {
+            _cbUpdater.BindTessControl(argument);
+        }
+
+        /// <summary>
+        /// Binds a uniform buffer for the tessellation evaluation shader stage.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void ConstantBufferBindTessEvaluation(int argument)
+        {
+            _cbUpdater.BindTessEvaluation(argument);
+        }
+
+        /// <summary>
+        /// Binds a uniform buffer for the geometry shader stage.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void ConstantBufferBindGeometry(int argument)
+        {
+            _cbUpdater.BindGeometry(argument);
+        }
+
+        /// <summary>
+        /// Binds a uniform buffer for the fragment shader stage.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void ConstantBufferBindFragment(int argument)
+        {
+            _cbUpdater.BindFragment(argument);
+        }
+
+        /// <summary>
+        /// Generic register read function that just returns 0.
+        /// </summary>
+        /// <returns>Zero</returns>
+        private static int Zero()
+        {
+            return 0;
+        }
+    }
+}

Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClassState.cs

@@ -0,0 +1,861 @@
+using Ryujinx.Common.Memory;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.InlineToMemory;
+using Ryujinx.Graphics.Gpu.Engine.Types;
+using Ryujinx.Graphics.Gpu.Image;
+using System;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Shader stage name.
+    /// </summary>
+    enum ShaderType
+    {
+        Vertex,
+        TessellationControl,
+        TessellationEvaluation,
+        Geometry,
+        Fragment
+    }
+
+    /// <summary>
+    /// Transform feedback buffer state.
+    /// </summary>
+    struct TfBufferState
+    {
+#pragma warning disable CS0649
+        public Boolean32 Enable;
+        public GpuVa Address;
+        public int Size;
+        public int Offset;
+        public uint Padding0;
+        public uint Padding1;
+        public uint Padding2;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Transform feedback state.
+    /// </summary>
+    struct TfState
+    {
+#pragma warning disable CS0649
+        public int BufferIndex;
+        public int VaryingsCount;
+        public int Stride;
+        public uint Padding;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Render target color buffer state.
+    /// </summary>
+    struct RtColorState
+    {
+#pragma warning disable CS0649
+        public GpuVa Address;
+        public int WidthOrStride;
+        public int Height;
+        public ColorFormat Format;
+        public MemoryLayout MemoryLayout;
+        public int Depth;
+        public int LayerSize;
+        public int BaseLayer;
+        public int Unknown0x24;
+        public int Padding0;
+        public int Padding1;
+        public int Padding2;
+        public int Padding3;
+        public int Padding4;
+        public int Padding5;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Viewport transform parameters, for viewport transformation.
+    /// </summary>
+    struct ViewportTransform
+    {
+#pragma warning disable CS0649
+        public float ScaleX;
+        public float ScaleY;
+        public float ScaleZ;
+        public float TranslateX;
+        public float TranslateY;
+        public float TranslateZ;
+        public uint Swizzle;
+        public uint SubpixelPrecisionBias;
+#pragma warning restore CS0649
+
+        /// <summary>
+        /// Unpacks viewport swizzle of the position X component.
+        /// </summary>
+        /// <returns>Swizzle enum value</returns>
+        public ViewportSwizzle UnpackSwizzleX()
+        {
+            return (ViewportSwizzle)(Swizzle & 7);
+        }
+
+        /// <summary>
+        /// Unpacks viewport swizzle of the position Y component.
+        /// </summary>
+        /// <returns>Swizzle enum value</returns>
+        public ViewportSwizzle UnpackSwizzleY()
+        {
+            return (ViewportSwizzle)((Swizzle >> 4) & 7);
+        }
+
+        /// <summary>
+        /// Unpacks viewport swizzle of the position Z component.
+        /// </summary>
+        /// <returns>Swizzle enum value</returns>
+        public ViewportSwizzle UnpackSwizzleZ()
+        {
+            return (ViewportSwizzle)((Swizzle >> 8) & 7);
+        }
+
+        /// <summary>
+        /// Unpacks viewport swizzle of the position W component.
+        /// </summary>
+        /// <returns>Swizzle enum value</returns>
+        public ViewportSwizzle UnpackSwizzleW()
+        {
+            return (ViewportSwizzle)((Swizzle >> 12) & 7);
+        }
+    }
+
+    /// <summary>
+    /// Viewport extents for viewport clipping, also includes depth range.
+    /// </summary>
+    struct ViewportExtents
+    {
+#pragma warning disable CS0649
+        public ushort X;
+        public ushort Width;
+        public ushort Y;
+        public ushort Height;
+        public float DepthNear;
+        public float DepthFar;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Draw state for non-indexed draws.
+    /// </summary>
+    struct VertexBufferDrawState
+    {
+#pragma warning disable CS0649
+        public int First;
+        public int Count;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Color buffer clear color.
+    /// </summary>
+    struct ClearColors
+    {
+#pragma warning disable CS0649
+        public float Red;
+        public float Green;
+        public float Blue;
+        public float Alpha;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Depth bias (also called polygon offset) parameters.
+    /// </summary>
+    struct DepthBiasState
+    {
+#pragma warning disable CS0649
+        public Boolean32 PointEnable;
+        public Boolean32 LineEnable;
+        public Boolean32 FillEnable;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Scissor state.
+    /// </summary>
+    struct ScissorState
+    {
+#pragma warning disable CS0649
+        public Boolean32 Enable;
+        public ushort X1;
+        public ushort X2;
+        public ushort Y1;
+        public ushort Y2;
+        public uint Padding;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Stencil test masks for back tests.
+    /// </summary>
+    struct StencilBackMasks
+    {
+#pragma warning disable CS0649
+        public int FuncRef;
+        public int Mask;
+        public int FuncMask;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Render target depth-stencil buffer state.
+    /// </summary>
+    struct RtDepthStencilState
+    {
+#pragma warning disable CS0649
+        public GpuVa Address;
+        public ZetaFormat Format;
+        public MemoryLayout MemoryLayout;
+        public int LayerSize;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Screen scissor state.
+    /// </summary>
+    struct ScreenScissorState
+    {
+#pragma warning disable CS0649
+        public ushort X;
+        public ushort Width;
+        public ushort Y;
+        public ushort Height;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Vertex buffer attribute state.
+    /// </summary>
+    struct VertexAttribState
+    {
+#pragma warning disable CS0649
+        public uint Attribute;
+#pragma warning restore CS0649
+
+        /// <summary>
+        /// Unpacks the index of the vertex buffer this attribute belongs to.
+        /// </summary>
+        /// <returns>Vertex buffer index</returns>
+        public int UnpackBufferIndex()
+        {
+            return (int)(Attribute & 0x1f);
+        }
+
+        /// <summary>
+        /// Unpacks the attribute constant flag.
+        /// </summary>
+        /// <returns>True if the attribute is constant, false otherwise</returns>
+        public bool UnpackIsConstant()
+        {
+            return (Attribute & 0x40) != 0;
+        }
+
+        /// <summary>
+        /// Unpacks the offset, in bytes, of the attribute on the vertex buffer.
+        /// </summary>
+        /// <returns>Attribute offset in bytes</returns>
+        public int UnpackOffset()
+        {
+            return (int)((Attribute >> 7) & 0x3fff);
+        }
+
+        /// <summary>
+        /// Unpacks the Maxwell attribute format integer.
+        /// </summary>
+        /// <returns>Attribute format integer</returns>
+        public uint UnpackFormat()
+        {
+            return Attribute & 0x3fe00000;
+        }
+    }
+
+    /// <summary>
+    /// Render target draw buffers control.
+    /// </summary>
+    struct RtControl
+    {
+#pragma warning disable CS0649
+        public uint Packed;
+#pragma warning restore CS0649
+
+        /// <summary>
+        /// Unpacks the number of active draw buffers.
+        /// </summary>
+        /// <returns>Number of active draw buffers</returns>
+        public int UnpackCount()
+        {
+            return (int)(Packed & 0xf);
+        }
+
+        /// <summary>
+        /// Unpacks the color attachment index for a given draw buffer.
+        /// </summary>
+        /// <param name="index">Index of the draw buffer</param>
+        /// <returns>Attachment index</returns>
+        public int UnpackPermutationIndex(int index)
+        {
+            return (int)((Packed >> (4 + index * 3)) & 7);
+        }
+    }
+
+    /// <summary>
+    /// 3D, 2D or 1D texture size.
+    /// </summary>
+    struct Size3D
+    {
+#pragma warning disable CS0649
+        public int Width;
+        public int Height;
+        public int Depth;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Stencil front test state and masks.
+    /// </summary>
+    struct StencilTestState
+    {
+#pragma warning disable CS0649
+        public Boolean32 Enable;
+        public StencilOp FrontSFail;
+        public StencilOp FrontDpFail;
+        public StencilOp FrontDpPass;
+        public CompareOp FrontFunc;
+        public int FrontFuncRef;
+        public int FrontFuncMask;
+        public int FrontMask;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Screen Y control register.
+    /// </summary>
+    [Flags]
+    enum YControl
+    {
+        NegateY = 1 << 0,
+        TriangleRastFlip = 1 << 4
+    }
+
+    /// <summary>
+    /// Condition for conditional rendering.
+    /// </summary>
+    enum Condition
+    {
+        Never,
+        Always,
+        ResultNonZero,
+        Equal,
+        NotEqual
+    }
+
+    /// <summary>
+    /// Texture or sampler pool state.
+    /// </summary>
+    struct PoolState
+    {
+#pragma warning disable CS0649
+        public GpuVa Address;
+        public int MaximumId;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Stencil back test state.
+    /// </summary>
+    struct StencilBackTestState
+    {
+#pragma warning disable CS0649
+        public Boolean32 TwoSided;
+        public StencilOp BackSFail;
+        public StencilOp BackDpFail;
+        public StencilOp BackDpPass;
+        public CompareOp BackFunc;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Primitive restart state.
+    /// </summary>
+    struct PrimitiveRestartState
+    {
+#pragma warning disable CS0649
+        public Boolean32 Enable;
+        public int Index;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// GPU index buffer state.
+    /// This is used on indexed draws.
+    /// </summary>
+    struct IndexBufferState
+    {
+#pragma warning disable CS0649
+        public GpuVa Address;
+        public GpuVa EndAddress;
+        public IndexType Type;
+        public int First;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Face culling and orientation parameters.
+    /// </summary>
+    struct FaceState
+    {
+#pragma warning disable CS0649
+        public Boolean32 CullEnable;
+        public FrontFace FrontFace;
+        public Face CullFace;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// View volume clip control.
+    /// </summary>
+    [Flags]
+    enum ViewVolumeClipControl
+    {
+        ForceDepthRangeZeroToOne = 1 << 0,
+        DepthClampDisabled = 1 << 11
+    }
+
+    /// <summary>
+    /// Logical operation state.
+    /// </summary>
+    struct LogicalOpState
+    {
+#pragma warning disable CS0649
+        public Boolean32 Enable;
+        public LogicalOp LogicalOp;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Render target color buffer mask.
+    /// This defines which color channels are written to the color buffer.
+    /// </summary>
+    struct RtColorMask
+    {
+#pragma warning disable CS0649
+        public uint Packed;
+#pragma warning restore CS0649
+
+        /// <summary>
+        /// Unpacks red channel enable.
+        /// </summary>
+        /// <returns>True to write the new red channel color, false to keep the old value</returns>
+        public bool UnpackRed()
+        {
+            return (Packed & 0x1) != 0;
+        }
+
+        /// <summary>
+        /// Unpacks green channel enable.
+        /// </summary>
+        /// <returns>True to write the new green channel color, false to keep the old value</returns>
+        public bool UnpackGreen()
+        {
+            return (Packed & 0x10) != 0;
+        }
+
+        /// <summary>
+        /// Unpacks blue channel enable.
+        /// </summary>
+        /// <returns>True to write the new blue channel color, false to keep the old value</returns>
+        public bool UnpackBlue()
+        {
+            return (Packed & 0x100) != 0;
+        }
+
+        /// <summary>
+        /// Unpacks alpha channel enable.
+        /// </summary>
+        /// <returns>True to write the new alpha channel color, false to keep the old value</returns>
+        public bool UnpackAlpha()
+        {
+            return (Packed & 0x1000) != 0;
+        }
+    }
+
+    /// <summary>
+    /// Vertex buffer state.
+    /// </summary>
+    struct VertexBufferState
+    {
+#pragma warning disable CS0649
+        public uint Control;
+        public GpuVa Address;
+        public int Divisor;
+#pragma warning restore CS0649
+
+        /// <summary>
+        /// Vertex buffer stride, defined as the number of bytes occupied by each vertex in memory.
+        /// </summary>
+        /// <returns>Vertex buffer stride</returns>
+        public int UnpackStride()
+        {
+            return (int)(Control & 0xfff);
+        }
+
+        /// <summary>
+        /// Vertex buffer enable.
+        /// </summary>
+        /// <returns>True if the vertex buffer is enabled, false otherwise</returns>
+        public bool UnpackEnable()
+        {
+            return (Control & (1 << 12)) != 0;
+        }
+    }
+
+    /// <summary>
+    /// Color buffer blending parameters, shared by all color buffers.
+    /// </summary>
+    struct BlendStateCommon
+    {
+#pragma warning disable CS0649
+        public Boolean32 SeparateAlpha;
+        public BlendOp ColorOp;
+        public BlendFactor ColorSrcFactor;
+        public BlendFactor ColorDstFactor;
+        public BlendOp AlphaOp;
+        public BlendFactor AlphaSrcFactor;
+        public uint Unknown0x1354;
+        public BlendFactor AlphaDstFactor;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Color buffer blending parameters.
+    /// </summary>
+    struct BlendState
+    {
+#pragma warning disable CS0649
+        public Boolean32 SeparateAlpha;
+        public BlendOp ColorOp;
+        public BlendFactor ColorSrcFactor;
+        public BlendFactor ColorDstFactor;
+        public BlendOp AlphaOp;
+        public BlendFactor AlphaSrcFactor;
+        public BlendFactor AlphaDstFactor;
+        public uint Padding;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// Graphics shader stage state.
+    /// </summary>
+    struct ShaderState
+    {
+#pragma warning disable CS0649
+        public uint Control;
+        public uint Offset;
+        public uint Unknown0x8;
+        public int MaxRegisters;
+        public ShaderType Type;
+        public uint Unknown0x14;
+        public uint Unknown0x18;
+        public uint Unknown0x1c;
+        public uint Unknown0x20;
+        public uint Unknown0x24;
+        public uint Unknown0x28;
+        public uint Unknown0x2c;
+        public uint Unknown0x30;
+        public uint Unknown0x34;
+        public uint Unknown0x38;
+        public uint Unknown0x3c;
+#pragma warning restore CS0649
+
+        /// <summary>
+        /// Unpacks shader enable information.
+        /// Must be ignored for vertex shaders, those are always enabled.
+        /// </summary>
+        /// <returns>True if the stage is enabled, false otherwise</returns>
+        public bool UnpackEnable()
+        {
+            return (Control & 1) != 0;
+        }
+    }
+
+    /// <summary>
+    /// Uniform buffer state for the uniform buffer currently being modified.
+    /// </summary>
+    struct UniformBufferState
+    {
+#pragma warning disable CS0649
+        public int Size;
+        public GpuVa Address;
+        public int Offset;
+#pragma warning restore CS0649
+    }
+
+    unsafe struct ThreedClassState : IShadowState
+    {
+#pragma warning disable CS0649
+        public uint SetObject;
+        public int SetObjectClassId => (int)((SetObject >> 0) & 0xFFFF);
+        public int SetObjectEngineId => (int)((SetObject >> 16) & 0x1F);
+        public fixed uint Reserved04[63];
+        public uint NoOperation;
+        public uint SetNotifyA;
+        public int SetNotifyAAddressUpper => (int)((SetNotifyA >> 0) & 0xFF);
+        public uint SetNotifyB;
+        public uint Notify;
+        public NotifyType NotifyType => (NotifyType)(Notify);
+        public uint WaitForIdle;
+        public uint LoadMmeInstructionRamPointer;
+        public uint LoadMmeInstructionRam;
+        public uint LoadMmeStartAddressRamPointer;
+        public uint LoadMmeStartAddressRam;
+        public uint SetMmeShadowRamControl;
+        public SetMmeShadowRamControlMode SetMmeShadowRamControlMode => (SetMmeShadowRamControlMode)((SetMmeShadowRamControl >> 0) & 0x3);
+        public fixed uint Reserved128[2];
+        public uint SetGlobalRenderEnableA;
+        public int SetGlobalRenderEnableAOffsetUpper => (int)((SetGlobalRenderEnableA >> 0) & 0xFF);
+        public uint SetGlobalRenderEnableB;
+        public uint SetGlobalRenderEnableC;
+        public int SetGlobalRenderEnableCMode => (int)((SetGlobalRenderEnableC >> 0) & 0x7);
+        public uint SendGoIdle;
+        public uint PmTrigger;
+        public uint PmTriggerWfi;
+        public fixed uint Reserved148[2];
+        public uint SetInstrumentationMethodHeader;
+        public uint SetInstrumentationMethodData;
+        public fixed uint Reserved158[10];
+        public uint LineLengthIn;
+        public uint LineCount;
+        public uint OffsetOutUpper;
+        public int OffsetOutUpperValue => (int)((OffsetOutUpper >> 0) & 0xFF);
+        public uint OffsetOut;
+        public uint PitchOut;
+        public uint SetDstBlockSize;
+        public SetDstBlockSizeWidth SetDstBlockSizeWidth => (SetDstBlockSizeWidth)((SetDstBlockSize >> 0) & 0xF);
+        public SetDstBlockSizeHeight SetDstBlockSizeHeight => (SetDstBlockSizeHeight)((SetDstBlockSize >> 4) & 0xF);
+        public SetDstBlockSizeDepth SetDstBlockSizeDepth => (SetDstBlockSizeDepth)((SetDstBlockSize >> 8) & 0xF);
+        public uint SetDstWidth;
+        public uint SetDstHeight;
+        public uint SetDstDepth;
+        public uint SetDstLayer;
+        public uint SetDstOriginBytesX;
+        public int SetDstOriginBytesXV => (int)((SetDstOriginBytesX >> 0) & 0xFFFFF);
+        public uint SetDstOriginSamplesY;
+        public int SetDstOriginSamplesYV => (int)((SetDstOriginSamplesY >> 0) & 0xFFFF);
+        public uint LaunchDma;
+        public LaunchDmaDstMemoryLayout LaunchDmaDstMemoryLayout => (LaunchDmaDstMemoryLayout)((LaunchDma >> 0) & 0x1);
+        public LaunchDmaCompletionType LaunchDmaCompletionType => (LaunchDmaCompletionType)((LaunchDma >> 4) & 0x3);
+        public LaunchDmaInterruptType LaunchDmaInterruptType => (LaunchDmaInterruptType)((LaunchDma >> 8) & 0x3);
+        public LaunchDmaSemaphoreStructSize LaunchDmaSemaphoreStructSize => (LaunchDmaSemaphoreStructSize)((LaunchDma >> 12) & 0x1);
+        public bool LaunchDmaReductionEnable => (LaunchDma & 0x2) != 0;
+        public LaunchDmaReductionOp LaunchDmaReductionOp => (LaunchDmaReductionOp)((LaunchDma >> 13) & 0x7);
+        public LaunchDmaReductionFormat LaunchDmaReductionFormat => (LaunchDmaReductionFormat)((LaunchDma >> 2) & 0x3);
+        public bool LaunchDmaSysmembarDisable => (LaunchDma & 0x40) != 0;
+        public uint LoadInlineData;
+        public fixed uint Reserved1B8[22];
+        public Boolean32 EarlyZForce;
+        public fixed uint Reserved214[45];
+        public uint SyncpointAction;
+        public fixed uint Reserved2CC[44];
+        public Boolean32 RasterizeEnable;
+        public Array4<TfBufferState> TfBufferState;
+        public fixed uint Reserved400[192];
+        public Array4<TfState> TfState;
+        public fixed uint Reserved740[1];
+        public Boolean32 TfEnable;
+        public fixed uint Reserved748[46];
+        public Array8<RtColorState> RtColorState;
+        public Array16<ViewportTransform> ViewportTransform;
+        public Array16<ViewportExtents> ViewportExtents;
+        public fixed uint ReservedD00[29];
+        public VertexBufferDrawState VertexBufferDrawState;
+        public uint DepthMode;
+        public ClearColors ClearColors;
+        public float ClearDepthValue;
+        public fixed uint ReservedD94[3];
+        public uint ClearStencilValue;
+        public fixed uint ReservedDA4[7];
+        public DepthBiasState DepthBiasState;
+        public fixed uint ReservedDCC[5];
+        public uint TextureBarrier;
+        public fixed uint ReservedDE4[7];
+        public Array16<ScissorState> ScissorState;
+        public fixed uint ReservedF00[21];
+        public StencilBackMasks StencilBackMasks;
+        public fixed uint ReservedF60[5];
+        public uint InvalidateTextures;
+        public fixed uint ReservedF78[1];
+        public uint TextureBarrierTiled;
+        public fixed uint ReservedF80[4];
+        public Boolean32 RtColorMaskShared;
+        public fixed uint ReservedF94[19];
+        public RtDepthStencilState RtDepthStencilState;
+        public ScreenScissorState ScreenScissorState;
+        public fixed uint ReservedFFC[89];
+        public Array16<VertexAttribState> VertexAttribState;
+        public fixed uint Reserved11A0[31];
+        public RtControl RtControl;
+        public fixed uint Reserved1220[2];
+        public Size3D RtDepthStencilSize;
+        public SamplerIndex SamplerIndex;
+        public fixed uint Reserved1238[37];
+        public Boolean32 DepthTestEnable;
+        public fixed uint Reserved12D0[5];
+        public Boolean32 BlendIndependent;
+        public Boolean32 DepthWriteEnable;
+        public Boolean32 AlphaTestEnable;
+        public fixed uint Reserved12F0[5];
+        public uint VbElementU8;
+        public uint Reserved1308;
+        public CompareOp DepthTestFunc;
+        public float AlphaTestRef;
+        public CompareOp AlphaTestFunc;
+        public uint Reserved1318;
+        public ColorF BlendConstant;
+        public fixed uint Reserved132C[4];
+        public BlendStateCommon BlendStateCommon;
+        public Boolean32 BlendEnableCommon;
+        public Array8<Boolean32> BlendEnable;
+        public StencilTestState StencilTestState;
+        public fixed uint Reserved13A0[3];
+        public YControl YControl;
+        public float LineWidthSmooth;
+        public float LineWidthAliased;
+        public fixed uint Reserved13B8[31];
+        public uint FirstVertex;
+        public uint FirstInstance;
+        public fixed uint Reserved143C[53];
+        public uint ClipDistanceEnable;
+        public uint Reserved1514;
+        public float PointSize;
+        public uint Reserved151C;
+        public Boolean32 PointSpriteEnable;
+        public fixed uint Reserved1524[3];
+        public uint ResetCounter;
+        public uint Reserved1534;
+        public Boolean32 RtDepthStencilEnable;
+        public fixed uint Reserved153C[5];
+        public GpuVa RenderEnableAddress;
+        public Condition RenderEnableCondition;
+        public PoolState SamplerPoolState;
+        public uint Reserved1568;
+        public float DepthBiasFactor;
+        public Boolean32 LineSmoothEnable;
+        public PoolState TexturePoolState;
+        public fixed uint Reserved1580[5];
+        public StencilBackTestState StencilBackTestState;
+        public fixed uint Reserved15A8[5];
+        public float DepthBiasUnits;
+        public fixed uint Reserved15C0[4];
+        public TextureMsaaMode RtMsaaMode;
+        public fixed uint Reserved15D4[5];
+        public uint VbElementU32;
+        public uint Reserved15EC;
+        public uint VbElementU16;
+        public fixed uint Reserved15F4[4];
+        public uint PointCoordReplace;
+        public GpuVa ShaderBaseAddress;
+        public uint Reserved1610;
+        public uint DrawEnd;
+        public uint DrawBegin;
+        public fixed uint Reserved161C[10];
+        public PrimitiveRestartState PrimitiveRestartState;
+        public fixed uint Reserved164C[95];
+        public IndexBufferState IndexBufferState;
+        public uint IndexBufferCount;
+        public uint DrawIndexedSmall;
+        public uint DrawIndexedSmall2;
+        public uint Reserved17EC;
+        public uint DrawIndexedSmallIncInstance;
+        public uint DrawIndexedSmallIncInstance2;
+        public fixed uint Reserved17F8[33];
+        public float DepthBiasClamp;
+        public Array16<Boolean32> VertexBufferInstanced;
+        public fixed uint Reserved18C0[20];
+        public Boolean32 VertexProgramPointSize;
+        public uint Reserved1914;
+        public FaceState FaceState;
+        public fixed uint Reserved1924[2];
+        public uint ViewportTransformEnable;
+        public fixed uint Reserved1930[3];
+        public ViewVolumeClipControl ViewVolumeClipControl;
+        public fixed uint Reserved1940[2];
+        public Boolean32 PrimitiveTypeOverrideEnable;
+        public fixed uint Reserved194C[9];
+        public PrimitiveTypeOverride PrimitiveTypeOverride;
+        public fixed uint Reserved1974[20];
+        public LogicalOpState LogicOpState;
+        public uint Reserved19CC;
+        public uint Clear;
+        public fixed uint Reserved19D4[11];
+        public Array8<RtColorMask> RtColorMask;
+        public fixed uint Reserved1A20[56];
+        public GpuVa SemaphoreAddress;
+        public int SemaphorePayload;
+        public uint SemaphoreControl;
+        public fixed uint Reserved1B10[60];
+        public Array16<VertexBufferState> VertexBufferState;
+        public fixed uint Reserved1D00[64];
+        public Array8<BlendState> BlendState;
+        public Array16<GpuVa> VertexBufferEndAddress;
+        public fixed uint Reserved1F80[32];
+        public Array6<ShaderState> ShaderState;
+        public fixed uint Reserved2180[96];
+        public uint SetFalcon00;
+        public uint SetFalcon01;
+        public uint SetFalcon02;
+        public uint SetFalcon03;
+        public uint SetFalcon04;
+        public uint SetFalcon05;
+        public uint SetFalcon06;
+        public uint SetFalcon07;
+        public uint SetFalcon08;
+        public uint SetFalcon09;
+        public uint SetFalcon10;
+        public uint SetFalcon11;
+        public uint SetFalcon12;
+        public uint SetFalcon13;
+        public uint SetFalcon14;
+        public uint SetFalcon15;
+        public uint SetFalcon16;
+        public uint SetFalcon17;
+        public uint SetFalcon18;
+        public uint SetFalcon19;
+        public uint SetFalcon20;
+        public uint SetFalcon21;
+        public uint SetFalcon22;
+        public uint SetFalcon23;
+        public uint SetFalcon24;
+        public uint SetFalcon25;
+        public uint SetFalcon26;
+        public uint SetFalcon27;
+        public uint SetFalcon28;
+        public uint SetFalcon29;
+        public uint SetFalcon30;
+        public uint SetFalcon31;
+        public UniformBufferState UniformBufferState;
+        public Array16<uint> UniformBufferUpdateData;
+        public fixed uint Reserved23D0[16];
+        public uint UniformBufferBindVertex;
+        public fixed uint Reserved2414[7];
+        public uint UniformBufferBindTessControl;
+        public fixed uint Reserved2434[7];
+        public uint UniformBufferBindTessEvaluation;
+        public fixed uint Reserved2454[7];
+        public uint UniformBufferBindGeometry;
+        public fixed uint Reserved2474[7];
+        public uint UniformBufferBindFragment;
+        public fixed uint Reserved2494[93];
+        public uint TextureBufferIndex;
+        public fixed uint Reserved260C[125];
+        public Array4<Array32<uint>> TfVaryingLocations;
+        public fixed uint Reserved2A00[640];
+        public MmeShadowScratch SetMmeShadowScratch;
+#pragma warning restore CS0649
+    }
+}