Move solution and projects to src

60 files changed, 17424 insertions, 0 deletions

diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Compute/ComputeClass.cs b/src/Ryujinx.Graphics.Gpu/Engine/Compute/ComputeClass.cs
new file mode 100644
index 00000000..2ac738fd
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Compute/ComputeClass.cs
@@ -0,0 +1,219 @@
+using Ryujinx.Graphics.Device;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.InlineToMemory;
+using Ryujinx.Graphics.Gpu.Engine.Threed;
+using Ryujinx.Graphics.Gpu.Engine.Types;
+using Ryujinx.Graphics.Gpu.Image;
+using Ryujinx.Graphics.Gpu.Shader;
+using Ryujinx.Graphics.Shader;
+using System;
+using System.Collections.Generic;
+using System.Runtime.CompilerServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Compute
+{
+    /// <summary>
+    /// Represents a compute engine class.
+    /// </summary>
+    class ComputeClass : IDeviceState
+    {
+        private readonly GpuContext _context;
+        private readonly GpuChannel _channel;
+        private readonly ThreedClass _3dEngine;
+        private readonly DeviceState<ComputeClassState> _state;
+
+        private readonly InlineToMemoryClass _i2mClass;
+
+        /// <summary>
+        /// Creates a new instance of the compute engine class.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        /// <param name="channel">GPU channel</param>
+        /// <param name="threedEngine">3D engine</param>
+        public ComputeClass(GpuContext context, GpuChannel channel, ThreedClass threedEngine)
+        {
+            _context = context;
+            _channel = channel;
+            _3dEngine = threedEngine;
+            _state = new DeviceState<ComputeClassState>(new Dictionary<string, RwCallback>
+            {
+                { nameof(ComputeClassState.LaunchDma), new RwCallback(LaunchDma, null) },
+                { nameof(ComputeClassState.LoadInlineData), new RwCallback(LoadInlineData, null) },
+                { nameof(ComputeClassState.SendSignalingPcasB), new RwCallback(SendSignalingPcasB, null) }
+            });
+
+            _i2mClass = new InlineToMemoryClass(context, channel, initializeState: false);
+        }
+
+        /// <summary>
+        /// Reads data from the class registers.
+        /// </summary>
+        /// <param name="offset">Register byte offset</param>
+        /// <returns>Data at the specified offset</returns>
+        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>
+        public void Write(int offset, int data) => _state.Write(offset, 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<ComputeClassState, InlineToMemoryClassState>(ref _state.State), argument);
+        }
+
+        /// <summary>
+        /// Pushes a block of data to the Inline-to-Memory engine.
+        /// </summary>
+        /// <param name="data">Data to push</param>
+        public void LoadInlineData(ReadOnlySpan<int> data)
+        {
+            _i2mClass.LoadInlineData(data);
+        }
+
+        /// <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 the compute dispatch operation.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void SendSignalingPcasB(int argument)
+        {
+            var memoryManager = _channel.MemoryManager;
+
+            // Since we're going to change the state, make sure any pending instanced draws are done.
+            _3dEngine.PerformDeferredDraws();
+
+            // Make sure all pending uniform buffer data is written to memory.
+            _3dEngine.FlushUboDirty();
+
+            uint qmdAddress = _state.State.SendPcasA;
+
+            var qmd = _channel.MemoryManager.Read<ComputeQmd>((ulong)qmdAddress << 8);
+
+            ulong shaderGpuVa = ((ulong)_state.State.SetProgramRegionAAddressUpper << 32) | _state.State.SetProgramRegionB;
+
+            shaderGpuVa += (uint)qmd.ProgramOffset;
+
+            int localMemorySize = qmd.ShaderLocalMemoryLowSize + qmd.ShaderLocalMemoryHighSize;
+
+            int sharedMemorySize = Math.Min(qmd.SharedMemorySize, _context.Capabilities.MaximumComputeSharedMemorySize);
+
+            for (int index = 0; index < Constants.TotalCpUniformBuffers; index++)
+            {
+                if (!qmd.ConstantBufferValid(index))
+                {
+                    continue;
+                }
+
+                ulong gpuVa = (uint)qmd.ConstantBufferAddrLower(index) | (ulong)qmd.ConstantBufferAddrUpper(index) << 32;
+                ulong size = (ulong)qmd.ConstantBufferSize(index);
+
+                _channel.BufferManager.SetComputeUniformBuffer(index, gpuVa, size);
+            }
+
+            ulong samplerPoolGpuVa = ((ulong)_state.State.SetTexSamplerPoolAOffsetUpper << 32) | _state.State.SetTexSamplerPoolB;
+            ulong texturePoolGpuVa = ((ulong)_state.State.SetTexHeaderPoolAOffsetUpper << 32) | _state.State.SetTexHeaderPoolB;
+
+            GpuChannelPoolState poolState = new GpuChannelPoolState(
+                texturePoolGpuVa,
+                _state.State.SetTexHeaderPoolCMaximumIndex,
+                _state.State.SetBindlessTextureConstantBufferSlotSelect);
+
+            GpuChannelComputeState computeState = new GpuChannelComputeState(
+                qmd.CtaThreadDimension0,
+                qmd.CtaThreadDimension1,
+                qmd.CtaThreadDimension2,
+                localMemorySize,
+                sharedMemorySize,
+                _channel.BufferManager.HasUnalignedStorageBuffers);
+
+            CachedShaderProgram cs = memoryManager.Physical.ShaderCache.GetComputeShader(_channel, poolState, computeState, shaderGpuVa);
+
+            _context.Renderer.Pipeline.SetProgram(cs.HostProgram);
+
+            _channel.TextureManager.SetComputeSamplerPool(samplerPoolGpuVa, _state.State.SetTexSamplerPoolCMaximumIndex, qmd.SamplerIndex);
+            _channel.TextureManager.SetComputeTexturePool(texturePoolGpuVa, _state.State.SetTexHeaderPoolCMaximumIndex);
+            _channel.TextureManager.SetComputeTextureBufferIndex(_state.State.SetBindlessTextureConstantBufferSlotSelect);
+
+            ShaderProgramInfo info = cs.Shaders[0].Info;
+
+            bool hasUnaligned = _channel.BufferManager.HasUnalignedStorageBuffers;
+
+            for (int index = 0; index < info.SBuffers.Count; index++)
+            {
+                BufferDescriptor sb = info.SBuffers[index];
+
+                ulong sbDescAddress = _channel.BufferManager.GetComputeUniformBufferAddress(0);
+
+                int sbDescOffset = 0x310 + sb.Slot * 0x10;
+
+                sbDescAddress += (ulong)sbDescOffset;
+
+                SbDescriptor sbDescriptor = _channel.MemoryManager.Physical.Read<SbDescriptor>(sbDescAddress);
+
+                _channel.BufferManager.SetComputeStorageBuffer(sb.Slot, sbDescriptor.PackAddress(), (uint)sbDescriptor.Size, sb.Flags);
+            }
+
+            if ((_channel.BufferManager.HasUnalignedStorageBuffers) != hasUnaligned)
+            {
+                // Refetch the shader, as assumptions about storage buffer alignment have changed.
+                cs = memoryManager.Physical.ShaderCache.GetComputeShader(_channel, poolState, computeState, shaderGpuVa);
+
+                _context.Renderer.Pipeline.SetProgram(cs.HostProgram);
+
+                info = cs.Shaders[0].Info;
+            }
+
+            for (int index = 0; index < info.CBuffers.Count; index++)
+            {
+                BufferDescriptor cb = info.CBuffers[index];
+
+                // NVN uses the "hardware" constant buffer for anything that is less than 8,
+                // and those are already bound above.
+                // Anything greater than or equal to 8 uses the emulated constant buffers.
+                // They are emulated using global memory loads.
+                if (cb.Slot < 8)
+                {
+                    continue;
+                }
+
+                ulong cbDescAddress = _channel.BufferManager.GetComputeUniformBufferAddress(0);
+
+                int cbDescOffset = 0x260 + (cb.Slot - 8) * 0x10;
+
+                cbDescAddress += (ulong)cbDescOffset;
+
+                SbDescriptor cbDescriptor = _channel.MemoryManager.Physical.Read<SbDescriptor>(cbDescAddress);
+
+                _channel.BufferManager.SetComputeUniformBuffer(cb.Slot, cbDescriptor.PackAddress(), (uint)cbDescriptor.Size);
+            }
+
+            _channel.BufferManager.SetComputeBufferBindings(cs.Bindings);
+
+            _channel.TextureManager.SetComputeBindings(cs.Bindings);
+
+            // Should never return false for mismatching spec state, since the shader was fetched above.
+            _channel.TextureManager.CommitComputeBindings(cs.SpecializationState);
+
+            _channel.BufferManager.CommitComputeBindings();
+
+            _context.Renderer.Pipeline.DispatchCompute(qmd.CtaRasterWidth, qmd.CtaRasterHeight, qmd.CtaRasterDepth);
+
+            _3dEngine.ForceShaderUpdate();
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Compute/ComputeClassState.cs b/src/Ryujinx.Graphics.Gpu/Engine/Compute/ComputeClassState.cs
new file mode 100644
index 00000000..5d81de5d
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Compute/ComputeClassState.cs
@@ -0,0 +1,435 @@
+// This file was auto-generated from NVIDIA official Maxwell definitions.
+
+using Ryujinx.Common.Memory;
+using Ryujinx.Graphics.Gpu.Engine.InlineToMemory;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Compute
+{
+    /// <summary>
+    /// Notify type.
+    /// </summary>
+    enum NotifyType
+    {
+        WriteOnly = 0,
+        WriteThenAwaken = 1,
+    }
+
+    /// <summary>
+    /// CWD control SM selection.
+    /// </summary>
+    enum SetCwdControlSmSelection
+    {
+        LoadBalanced = 0,
+        RoundRobin = 1,
+    }
+
+    /// <summary>
+    /// Cache lines to invalidate.
+    /// </summary>
+    enum InvalidateCacheLines
+    {
+        All = 0,
+        One = 1,
+    }
+
+    /// <summary>
+    /// GWC SCG type.
+    /// </summary>
+    enum SetGwcScgTypeScgType
+    {
+        GraphicsCompute0 = 0,
+        Compute1 = 1,
+    }
+
+    /// <summary>
+    /// Render enable override mode.
+    /// </summary>
+    enum SetRenderEnableOverrideMode
+    {
+        UseRenderEnable = 0,
+        AlwaysRender = 1,
+        NeverRender = 2,
+    }
+
+    /// <summary>
+    /// Semaphore report operation.
+    /// </summary>
+    enum SetReportSemaphoreDOperation
+    {
+        Release = 0,
+        Trap = 3,
+    }
+
+    /// <summary>
+    /// Semaphore report structure size.
+    /// </summary>
+    enum SetReportSemaphoreDStructureSize
+    {
+        FourWords = 0,
+        OneWord = 1,
+    }
+
+    /// <summary>
+    /// Semaphore report reduction operation.
+    /// </summary>
+    enum SetReportSemaphoreDReductionOp
+    {
+        RedAdd = 0,
+        RedMin = 1,
+        RedMax = 2,
+        RedInc = 3,
+        RedDec = 4,
+        RedAnd = 5,
+        RedOr = 6,
+        RedXor = 7,
+    }
+
+    /// <summary>
+    /// Semaphore report reduction format.
+    /// </summary>
+    enum SetReportSemaphoreDReductionFormat
+    {
+        Unsigned32 = 0,
+        Signed32 = 1,
+    }
+
+    /// <summary>
+    /// Compute class state.
+    /// </summary>
+    unsafe struct ComputeClassState
+    {
+#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 fixed uint Reserved114[7];
+        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[9];
+        public uint SetI2mSemaphoreA;
+        public int SetI2mSemaphoreAOffsetUpper => (int)((SetI2mSemaphoreA >> 0) & 0xFF);
+        public uint SetI2mSemaphoreB;
+        public uint SetI2mSemaphoreC;
+        public fixed uint Reserved1E8[2];
+        public uint SetI2mSpareNoop00;
+        public uint SetI2mSpareNoop01;
+        public uint SetI2mSpareNoop02;
+        public uint SetI2mSpareNoop03;
+        public uint SetValidSpanOverflowAreaA;
+        public int SetValidSpanOverflowAreaAAddressUpper => (int)((SetValidSpanOverflowAreaA >> 0) & 0xFF);
+        public uint SetValidSpanOverflowAreaB;
+        public uint SetValidSpanOverflowAreaC;
+        public uint SetCoalesceWaitingPeriodUnit;
+        public uint PerfmonTransfer;
+        public uint SetShaderSharedMemoryWindow;
+        public uint SetSelectMaxwellTextureHeaders;
+        public bool SetSelectMaxwellTextureHeadersV => (SetSelectMaxwellTextureHeaders & 0x1) != 0;
+        public uint InvalidateShaderCaches;
+        public bool InvalidateShaderCachesInstruction => (InvalidateShaderCaches & 0x1) != 0;
+        public bool InvalidateShaderCachesData => (InvalidateShaderCaches & 0x10) != 0;
+        public bool InvalidateShaderCachesConstant => (InvalidateShaderCaches & 0x1000) != 0;
+        public bool InvalidateShaderCachesLocks => (InvalidateShaderCaches & 0x2) != 0;
+        public bool InvalidateShaderCachesFlushData => (InvalidateShaderCaches & 0x4) != 0;
+        public uint SetReservedSwMethod00;
+        public uint SetReservedSwMethod01;
+        public uint SetReservedSwMethod02;
+        public uint SetReservedSwMethod03;
+        public uint SetReservedSwMethod04;
+        public uint SetReservedSwMethod05;
+        public uint SetReservedSwMethod06;
+        public uint SetReservedSwMethod07;
+        public uint SetCwdControl;
+        public SetCwdControlSmSelection SetCwdControlSmSelection => (SetCwdControlSmSelection)((SetCwdControl >> 0) & 0x1);
+        public uint InvalidateTextureHeaderCacheNoWfi;
+        public InvalidateCacheLines InvalidateTextureHeaderCacheNoWfiLines => (InvalidateCacheLines)((InvalidateTextureHeaderCacheNoWfi >> 0) & 0x1);
+        public int InvalidateTextureHeaderCacheNoWfiTag => (int)((InvalidateTextureHeaderCacheNoWfi >> 4) & 0x3FFFFF);
+        public uint SetCwdRefCounter;
+        public int SetCwdRefCounterSelect => (int)((SetCwdRefCounter >> 0) & 0x3F);
+        public int SetCwdRefCounterValue => (int)((SetCwdRefCounter >> 8) & 0xFFFF);
+        public uint SetReservedSwMethod08;
+        public uint SetReservedSwMethod09;
+        public uint SetReservedSwMethod10;
+        public uint SetReservedSwMethod11;
+        public uint SetReservedSwMethod12;
+        public uint SetReservedSwMethod13;
+        public uint SetReservedSwMethod14;
+        public uint SetReservedSwMethod15;
+        public uint SetGwcScgType;
+        public SetGwcScgTypeScgType SetGwcScgTypeScgType => (SetGwcScgTypeScgType)((SetGwcScgType >> 0) & 0x1);
+        public uint SetScgControl;
+        public int SetScgControlCompute1MaxSmCount => (int)((SetScgControl >> 0) & 0x1FF);
+        public uint InvalidateConstantBufferCacheA;
+        public int InvalidateConstantBufferCacheAAddressUpper => (int)((InvalidateConstantBufferCacheA >> 0) & 0xFF);
+        public uint InvalidateConstantBufferCacheB;
+        public uint InvalidateConstantBufferCacheC;
+        public int InvalidateConstantBufferCacheCByteCount => (int)((InvalidateConstantBufferCacheC >> 0) & 0x1FFFF);
+        public bool InvalidateConstantBufferCacheCThruL2 => (InvalidateConstantBufferCacheC & 0x80000000) != 0;
+        public uint SetComputeClassVersion;
+        public int SetComputeClassVersionCurrent => (int)((SetComputeClassVersion >> 0) & 0xFFFF);
+        public int SetComputeClassVersionOldestSupported => (int)((SetComputeClassVersion >> 16) & 0xFFFF);
+        public uint CheckComputeClassVersion;
+        public int CheckComputeClassVersionCurrent => (int)((CheckComputeClassVersion >> 0) & 0xFFFF);
+        public int CheckComputeClassVersionOldestSupported => (int)((CheckComputeClassVersion >> 16) & 0xFFFF);
+        public uint SetQmdVersion;
+        public int SetQmdVersionCurrent => (int)((SetQmdVersion >> 0) & 0xFFFF);
+        public int SetQmdVersionOldestSupported => (int)((SetQmdVersion >> 16) & 0xFFFF);
+        public uint SetWfiConfig;
+        public bool SetWfiConfigEnableScgTypeWfi => (SetWfiConfig & 0x1) != 0;
+        public uint CheckQmdVersion;
+        public int CheckQmdVersionCurrent => (int)((CheckQmdVersion >> 0) & 0xFFFF);
+        public int CheckQmdVersionOldestSupported => (int)((CheckQmdVersion >> 16) & 0xFFFF);
+        public uint WaitForIdleScgType;
+        public uint InvalidateSkedCaches;
+        public bool InvalidateSkedCachesV => (InvalidateSkedCaches & 0x1) != 0;
+        public uint SetScgRenderEnableControl;
+        public bool SetScgRenderEnableControlCompute1UsesRenderEnable => (SetScgRenderEnableControl & 0x1) != 0;
+        public fixed uint Reserved2A0[4];
+        public uint SetCwdSlotCount;
+        public int SetCwdSlotCountV => (int)((SetCwdSlotCount >> 0) & 0xFF);
+        public uint SendPcasA;
+        public uint SendPcasB;
+        public int SendPcasBFrom => (int)((SendPcasB >> 0) & 0xFFFFFF);
+        public int SendPcasBDelta => (int)((SendPcasB >> 24) & 0xFF);
+        public uint SendSignalingPcasB;
+        public bool SendSignalingPcasBInvalidate => (SendSignalingPcasB & 0x1) != 0;
+        public bool SendSignalingPcasBSchedule => (SendSignalingPcasB & 0x2) != 0;
+        public fixed uint Reserved2C0[9];
+        public uint SetShaderLocalMemoryNonThrottledA;
+        public int SetShaderLocalMemoryNonThrottledASizeUpper => (int)((SetShaderLocalMemoryNonThrottledA >> 0) & 0xFF);
+        public uint SetShaderLocalMemoryNonThrottledB;
+        public uint SetShaderLocalMemoryNonThrottledC;
+        public int SetShaderLocalMemoryNonThrottledCMaxSmCount => (int)((SetShaderLocalMemoryNonThrottledC >> 0) & 0x1FF);
+        public uint SetShaderLocalMemoryThrottledA;
+        public int SetShaderLocalMemoryThrottledASizeUpper => (int)((SetShaderLocalMemoryThrottledA >> 0) & 0xFF);
+        public uint SetShaderLocalMemoryThrottledB;
+        public uint SetShaderLocalMemoryThrottledC;
+        public int SetShaderLocalMemoryThrottledCMaxSmCount => (int)((SetShaderLocalMemoryThrottledC >> 0) & 0x1FF);
+        public fixed uint Reserved2FC[5];
+        public uint SetSpaVersion;
+        public int SetSpaVersionMinor => (int)((SetSpaVersion >> 0) & 0xFF);
+        public int SetSpaVersionMajor => (int)((SetSpaVersion >> 8) & 0xFF);
+        public fixed uint Reserved314[123];
+        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 fixed uint Reserved580[127];
+        public uint SetShaderLocalMemoryWindow;
+        public fixed uint Reserved780[4];
+        public uint SetShaderLocalMemoryA;
+        public int SetShaderLocalMemoryAAddressUpper => (int)((SetShaderLocalMemoryA >> 0) & 0xFF);
+        public uint SetShaderLocalMemoryB;
+        public fixed uint Reserved798[383];
+        public uint SetShaderCacheControl;
+        public bool SetShaderCacheControlIcachePrefetchEnable => (SetShaderCacheControl & 0x1) != 0;
+        public fixed uint ReservedD98[19];
+        public uint SetSmTimeoutInterval;
+        public int SetSmTimeoutIntervalCounterBit => (int)((SetSmTimeoutInterval >> 0) & 0x3F);
+        public fixed uint ReservedDE8[87];
+        public uint SetSpareNoop12;
+        public uint SetSpareNoop13;
+        public uint SetSpareNoop14;
+        public uint SetSpareNoop15;
+        public fixed uint ReservedF54[59];
+        public uint SetSpareNoop00;
+        public uint SetSpareNoop01;
+        public uint SetSpareNoop02;
+        public uint SetSpareNoop03;
+        public uint SetSpareNoop04;
+        public uint SetSpareNoop05;
+        public uint SetSpareNoop06;
+        public uint SetSpareNoop07;
+        public uint SetSpareNoop08;
+        public uint SetSpareNoop09;
+        public uint SetSpareNoop10;
+        public uint SetSpareNoop11;
+        public fixed uint Reserved1070[103];
+        public uint InvalidateSamplerCacheAll;
+        public bool InvalidateSamplerCacheAllV => (InvalidateSamplerCacheAll & 0x1) != 0;
+        public uint InvalidateTextureHeaderCacheAll;
+        public bool InvalidateTextureHeaderCacheAllV => (InvalidateTextureHeaderCacheAll & 0x1) != 0;
+        public fixed uint Reserved1214[29];
+        public uint InvalidateTextureDataCacheNoWfi;
+        public InvalidateCacheLines InvalidateTextureDataCacheNoWfiLines => (InvalidateCacheLines)((InvalidateTextureDataCacheNoWfi >> 0) & 0x1);
+        public int InvalidateTextureDataCacheNoWfiTag => (int)((InvalidateTextureDataCacheNoWfi >> 4) & 0x3FFFFF);
+        public fixed uint Reserved128C[7];
+        public uint ActivatePerfSettingsForComputeContext;
+        public bool ActivatePerfSettingsForComputeContextAll => (ActivatePerfSettingsForComputeContext & 0x1) != 0;
+        public fixed uint Reserved12AC[33];
+        public uint InvalidateSamplerCache;
+        public InvalidateCacheLines InvalidateSamplerCacheLines => (InvalidateCacheLines)((InvalidateSamplerCache >> 0) & 0x1);
+        public int InvalidateSamplerCacheTag => (int)((InvalidateSamplerCache >> 4) & 0x3FFFFF);
+        public uint InvalidateTextureHeaderCache;
+        public InvalidateCacheLines InvalidateTextureHeaderCacheLines => (InvalidateCacheLines)((InvalidateTextureHeaderCache >> 0) & 0x1);
+        public int InvalidateTextureHeaderCacheTag => (int)((InvalidateTextureHeaderCache >> 4) & 0x3FFFFF);
+        public uint InvalidateTextureDataCache;
+        public InvalidateCacheLines InvalidateTextureDataCacheLines => (InvalidateCacheLines)((InvalidateTextureDataCache >> 0) & 0x1);
+        public int InvalidateTextureDataCacheTag => (int)((InvalidateTextureDataCache >> 4) & 0x3FFFFF);
+        public fixed uint Reserved133C[58];
+        public uint InvalidateSamplerCacheNoWfi;
+        public InvalidateCacheLines InvalidateSamplerCacheNoWfiLines => (InvalidateCacheLines)((InvalidateSamplerCacheNoWfi >> 0) & 0x1);
+        public int InvalidateSamplerCacheNoWfiTag => (int)((InvalidateSamplerCacheNoWfi >> 4) & 0x3FFFFF);
+        public fixed uint Reserved1428[64];
+        public uint SetShaderExceptions;
+        public bool SetShaderExceptionsEnable => (SetShaderExceptions & 0x1) != 0;
+        public fixed uint Reserved152C[9];
+        public uint SetRenderEnableA;
+        public int SetRenderEnableAOffsetUpper => (int)((SetRenderEnableA >> 0) & 0xFF);
+        public uint SetRenderEnableB;
+        public uint SetRenderEnableC;
+        public int SetRenderEnableCMode => (int)((SetRenderEnableC >> 0) & 0x7);
+        public uint SetTexSamplerPoolA;
+        public int SetTexSamplerPoolAOffsetUpper => (int)((SetTexSamplerPoolA >> 0) & 0xFF);
+        public uint SetTexSamplerPoolB;
+        public uint SetTexSamplerPoolC;
+        public int SetTexSamplerPoolCMaximumIndex => (int)((SetTexSamplerPoolC >> 0) & 0xFFFFF);
+        public fixed uint Reserved1568[3];
+        public uint SetTexHeaderPoolA;
+        public int SetTexHeaderPoolAOffsetUpper => (int)((SetTexHeaderPoolA >> 0) & 0xFF);
+        public uint SetTexHeaderPoolB;
+        public uint SetTexHeaderPoolC;
+        public int SetTexHeaderPoolCMaximumIndex => (int)((SetTexHeaderPoolC >> 0) & 0x3FFFFF);
+        public fixed uint Reserved1580[34];
+        public uint SetProgramRegionA;
+        public int SetProgramRegionAAddressUpper => (int)((SetProgramRegionA >> 0) & 0xFF);
+        public uint SetProgramRegionB;
+        public fixed uint Reserved1610[34];
+        public uint InvalidateShaderCachesNoWfi;
+        public bool InvalidateShaderCachesNoWfiInstruction => (InvalidateShaderCachesNoWfi & 0x1) != 0;
+        public bool InvalidateShaderCachesNoWfiGlobalData => (InvalidateShaderCachesNoWfi & 0x10) != 0;
+        public bool InvalidateShaderCachesNoWfiConstant => (InvalidateShaderCachesNoWfi & 0x1000) != 0;
+        public fixed uint Reserved169C[170];
+        public uint SetRenderEnableOverride;
+        public SetRenderEnableOverrideMode SetRenderEnableOverrideMode => (SetRenderEnableOverrideMode)((SetRenderEnableOverride >> 0) & 0x3);
+        public fixed uint Reserved1948[57];
+        public uint PipeNop;
+        public uint SetSpare00;
+        public uint SetSpare01;
+        public uint SetSpare02;
+        public uint SetSpare03;
+        public fixed uint Reserved1A40[48];
+        public uint SetReportSemaphoreA;
+        public int SetReportSemaphoreAOffsetUpper => (int)((SetReportSemaphoreA >> 0) & 0xFF);
+        public uint SetReportSemaphoreB;
+        public uint SetReportSemaphoreC;
+        public uint SetReportSemaphoreD;
+        public SetReportSemaphoreDOperation SetReportSemaphoreDOperation => (SetReportSemaphoreDOperation)((SetReportSemaphoreD >> 0) & 0x3);
+        public bool SetReportSemaphoreDAwakenEnable => (SetReportSemaphoreD & 0x100000) != 0;
+        public SetReportSemaphoreDStructureSize SetReportSemaphoreDStructureSize => (SetReportSemaphoreDStructureSize)((SetReportSemaphoreD >> 28) & 0x1);
+        public bool SetReportSemaphoreDFlushDisable => (SetReportSemaphoreD & 0x4) != 0;
+        public bool SetReportSemaphoreDReductionEnable => (SetReportSemaphoreD & 0x8) != 0;
+        public SetReportSemaphoreDReductionOp SetReportSemaphoreDReductionOp => (SetReportSemaphoreDReductionOp)((SetReportSemaphoreD >> 9) & 0x7);
+        public SetReportSemaphoreDReductionFormat SetReportSemaphoreDReductionFormat => (SetReportSemaphoreDReductionFormat)((SetReportSemaphoreD >> 17) & 0x3);
+        public fixed uint Reserved1B10[702];
+        public uint SetBindlessTexture;
+        public int SetBindlessTextureConstantBufferSlotSelect => (int)((SetBindlessTexture >> 0) & 0x7);
+        public uint SetTrapHandler;
+        public fixed uint Reserved2610[843];
+        public Array8<uint> SetShaderPerformanceCounterValueUpper;
+        public Array8<uint> SetShaderPerformanceCounterValue;
+        public Array8<uint> SetShaderPerformanceCounterEvent;
+        public int SetShaderPerformanceCounterEventEvent(int i) => (int)((SetShaderPerformanceCounterEvent[i] >> 0) & 0xFF);
+        public Array8<uint> SetShaderPerformanceCounterControlA;
+        public int SetShaderPerformanceCounterControlAEvent0(int i) => (int)((SetShaderPerformanceCounterControlA[i] >> 0) & 0x3);
+        public int SetShaderPerformanceCounterControlABitSelect0(int i) => (int)((SetShaderPerformanceCounterControlA[i] >> 2) & 0x7);
+        public int SetShaderPerformanceCounterControlAEvent1(int i) => (int)((SetShaderPerformanceCounterControlA[i] >> 5) & 0x3);
+        public int SetShaderPerformanceCounterControlABitSelect1(int i) => (int)((SetShaderPerformanceCounterControlA[i] >> 7) & 0x7);
+        public int SetShaderPerformanceCounterControlAEvent2(int i) => (int)((SetShaderPerformanceCounterControlA[i] >> 10) & 0x3);
+        public int SetShaderPerformanceCounterControlABitSelect2(int i) => (int)((SetShaderPerformanceCounterControlA[i] >> 12) & 0x7);
+        public int SetShaderPerformanceCounterControlAEvent3(int i) => (int)((SetShaderPerformanceCounterControlA[i] >> 15) & 0x3);
+        public int SetShaderPerformanceCounterControlABitSelect3(int i) => (int)((SetShaderPerformanceCounterControlA[i] >> 17) & 0x7);
+        public int SetShaderPerformanceCounterControlAEvent4(int i) => (int)((SetShaderPerformanceCounterControlA[i] >> 20) & 0x3);
+        public int SetShaderPerformanceCounterControlABitSelect4(int i) => (int)((SetShaderPerformanceCounterControlA[i] >> 22) & 0x7);
+        public int SetShaderPerformanceCounterControlAEvent5(int i) => (int)((SetShaderPerformanceCounterControlA[i] >> 25) & 0x3);
+        public int SetShaderPerformanceCounterControlABitSelect5(int i) => (int)((SetShaderPerformanceCounterControlA[i] >> 27) & 0x7);
+        public int SetShaderPerformanceCounterControlASpare(int i) => (int)((SetShaderPerformanceCounterControlA[i] >> 30) & 0x3);
+        public Array8<uint> SetShaderPerformanceCounterControlB;
+        public bool SetShaderPerformanceCounterControlBEdge(int i) => (SetShaderPerformanceCounterControlB[i] & 0x1) != 0;
+        public int SetShaderPerformanceCounterControlBMode(int i) => (int)((SetShaderPerformanceCounterControlB[i] >> 1) & 0x3);
+        public bool SetShaderPerformanceCounterControlBWindowed(int i) => (SetShaderPerformanceCounterControlB[i] & 0x8) != 0;
+        public int SetShaderPerformanceCounterControlBFunc(int i) => (int)((SetShaderPerformanceCounterControlB[i] >> 4) & 0xFFFF);
+        public uint SetShaderPerformanceCounterTrapControl;
+        public int SetShaderPerformanceCounterTrapControlMask => (int)((SetShaderPerformanceCounterTrapControl >> 0) & 0xFF);
+        public uint StartShaderPerformanceCounter;
+        public int StartShaderPerformanceCounterCounterMask => (int)((StartShaderPerformanceCounter >> 0) & 0xFF);
+        public uint StopShaderPerformanceCounter;
+        public int StopShaderPerformanceCounterCounterMask => (int)((StopShaderPerformanceCounter >> 0) & 0xFF);
+        public fixed uint Reserved33E8[6];
+        public MmeShadowScratch SetMmeShadowScratch;
+#pragma warning restore CS0649
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Compute/ComputeQmd.cs b/src/Ryujinx.Graphics.Gpu/Engine/Compute/ComputeQmd.cs
new file mode 100644
index 00000000..1b20e41c
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Compute/ComputeQmd.cs
@@ -0,0 +1,275 @@
+using Ryujinx.Graphics.Gpu.Engine.Types;
+using System;
+using System.Runtime.CompilerServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Compute
+{
+    /// <summary>
+    /// Type of the dependent Queue Meta Data.
+    /// </summary>
+    enum DependentQmdType
+    {
+        Queue,
+        Grid
+    }
+
+    /// <summary>
+    /// Type of the release memory barrier.
+    /// </summary>
+    enum ReleaseMembarType
+    {
+        FeNone,
+        FeSysmembar
+    }
+
+    /// <summary>
+    /// Type of the CWD memory barrier.
+    /// </summary>
+    enum CwdMembarType
+    {
+        L1None,
+        L1Sysmembar,
+        L1Membar
+    }
+
+    /// <summary>
+    /// NaN behavior of 32-bits float operations on the shader.
+    /// </summary>
+    enum Fp32NanBehavior
+    {
+        Legacy,
+        Fp64Compatible
+    }
+
+    /// <summary>
+    /// NaN behavior of 32-bits float to integer conversion on the shader.
+    /// </summary>
+    enum Fp32F2iNanBehavior
+    {
+        PassZero,
+        PassIndefinite
+    }
+
+    /// <summary>
+    /// Limit of calls.
+    /// </summary>
+    enum ApiVisibleCallLimit
+    {
+        _32,
+        NoCheck
+    }
+
+    /// <summary>
+    /// Shared memory bank mapping mode.
+    /// </summary>
+    enum SharedMemoryBankMapping
+    {
+        FourBytesPerBank,
+        EightBytesPerBank
+    }
+
+    /// <summary>
+    /// Denormal behavior of 32-bits float narrowing instructions.
+    /// </summary>
+    enum Fp32NarrowInstruction
+    {
+        KeepDenorms,
+        FlushDenorms
+    }
+
+    /// <summary>
+    /// Configuration of the L1 cache.
+    /// </summary>
+    enum L1Configuration
+    {
+        DirectlyAddressableMemorySize16kb,
+        DirectlyAddressableMemorySize32kb,
+        DirectlyAddressableMemorySize48kb
+    }
+
+    /// <summary>
+    /// Reduction operation.
+    /// </summary>
+    enum ReductionOp
+    {
+        RedAdd,
+        RedMin,
+        RedMax,
+        RedInc,
+        RedDec,
+        RedAnd,
+        RedOr,
+        RedXor
+    }
+
+    /// <summary>
+    /// Reduction format.
+    /// </summary>
+    enum ReductionFormat
+    {
+        Unsigned32,
+        Signed32
+    }
+
+    /// <summary>
+    /// Size of a structure in words.
+    /// </summary>
+    enum StructureSize
+    {
+        FourWords,
+        OneWord
+    }
+
+    /// <summary>
+    /// Compute Queue Meta Data.
+    /// </summary>
+    unsafe struct ComputeQmd
+    {
+        private fixed int _words[64];
+
+        public int OuterPut => BitRange(30, 0);
+        public bool OuterOverflow => Bit(31);
+        public int OuterGet => BitRange(62, 32);
+        public bool OuterStickyOverflow => Bit(63);
+        public int InnerGet => BitRange(94, 64);
+        public bool InnerOverflow => Bit(95);
+        public int InnerPut => BitRange(126, 96);
+        public bool InnerStickyOverflow => Bit(127);
+        public int QmdReservedAA => BitRange(159, 128);
+        public int DependentQmdPointer => BitRange(191, 160);
+        public int QmdGroupId => BitRange(197, 192);
+        public bool SmGlobalCachingEnable => Bit(198);
+        public bool RunCtaInOneSmPartition => Bit(199);
+        public bool IsQueue => Bit(200);
+        public bool AddToHeadOfQmdGroupLinkedList => Bit(201);
+        public bool SemaphoreReleaseEnable0 => Bit(202);
+        public bool SemaphoreReleaseEnable1 => Bit(203);
+        public bool RequireSchedulingPcas => Bit(204);
+        public bool DependentQmdScheduleEnable => Bit(205);
+        public DependentQmdType DependentQmdType => (DependentQmdType)BitRange(206, 206);
+        public bool DependentQmdFieldCopy => Bit(207);
+        public int QmdReservedB => BitRange(223, 208);
+        public int CircularQueueSize => BitRange(248, 224);
+        public bool QmdReservedC => Bit(249);
+        public bool InvalidateTextureHeaderCache => Bit(250);
+        public bool InvalidateTextureSamplerCache => Bit(251);
+        public bool InvalidateTextureDataCache => Bit(252);
+        public bool InvalidateShaderDataCache => Bit(253);
+        public bool InvalidateInstructionCache => Bit(254);
+        public bool InvalidateShaderConstantCache => Bit(255);
+        public int ProgramOffset => BitRange(287, 256);
+        public int CircularQueueAddrLower => BitRange(319, 288);
+        public int CircularQueueAddrUpper => BitRange(327, 320);
+        public int QmdReservedD => BitRange(335, 328);
+        public int CircularQueueEntrySize => BitRange(351, 336);
+        public int CwdReferenceCountId => BitRange(357, 352);
+        public int CwdReferenceCountDeltaMinusOne => BitRange(365, 358);
+        public ReleaseMembarType ReleaseMembarType => (ReleaseMembarType)BitRange(366, 366);
+        public bool CwdReferenceCountIncrEnable => Bit(367);
+        public CwdMembarType CwdMembarType => (CwdMembarType)BitRange(369, 368);
+        public bool SequentiallyRunCtas => Bit(370);
+        public bool CwdReferenceCountDecrEnable => Bit(371);
+        public bool Throttled => Bit(372);
+        public Fp32NanBehavior Fp32NanBehavior => (Fp32NanBehavior)BitRange(376, 376);
+        public Fp32F2iNanBehavior Fp32F2iNanBehavior => (Fp32F2iNanBehavior)BitRange(377, 377);
+        public ApiVisibleCallLimit ApiVisibleCallLimit => (ApiVisibleCallLimit)BitRange(378, 378);
+        public SharedMemoryBankMapping SharedMemoryBankMapping => (SharedMemoryBankMapping)BitRange(379, 379);
+        public SamplerIndex SamplerIndex => (SamplerIndex)BitRange(382, 382);
+        public Fp32NarrowInstruction Fp32NarrowInstruction => (Fp32NarrowInstruction)BitRange(383, 383);
+        public int CtaRasterWidth => BitRange(415, 384);
+        public int CtaRasterHeight => BitRange(431, 416);
+        public int CtaRasterDepth => BitRange(447, 432);
+        public int CtaRasterWidthResume => BitRange(479, 448);
+        public int CtaRasterHeightResume => BitRange(495, 480);
+        public int CtaRasterDepthResume => BitRange(511, 496);
+        public int QueueEntriesPerCtaMinusOne => BitRange(518, 512);
+        public int CoalesceWaitingPeriod => BitRange(529, 522);
+        public int SharedMemorySize => BitRange(561, 544);
+        public int QmdReservedG => BitRange(575, 562);
+        public int QmdVersion => BitRange(579, 576);
+        public int QmdMajorVersion => BitRange(583, 580);
+        public int QmdReservedH => BitRange(591, 584);
+        public int CtaThreadDimension0 => BitRange(607, 592);
+        public int CtaThreadDimension1 => BitRange(623, 608);
+        public int CtaThreadDimension2 => BitRange(639, 624);
+        public bool ConstantBufferValid(int i) => Bit(640 + i * 1);
+        public int QmdReservedI => BitRange(668, 648);
+        public L1Configuration L1Configuration => (L1Configuration)BitRange(671, 669);
+        public int SmDisableMaskLower => BitRange(703, 672);
+        public int SmDisableMaskUpper => BitRange(735, 704);
+        public int Release0AddressLower => BitRange(767, 736);
+        public int Release0AddressUpper => BitRange(775, 768);
+        public int QmdReservedJ => BitRange(783, 776);
+        public ReductionOp Release0ReductionOp => (ReductionOp)BitRange(790, 788);
+        public bool QmdReservedK => Bit(791);
+        public ReductionFormat Release0ReductionFormat => (ReductionFormat)BitRange(793, 792);
+        public bool Release0ReductionEnable => Bit(794);
+        public StructureSize Release0StructureSize => (StructureSize)BitRange(799, 799);
+        public int Release0Payload => BitRange(831, 800);
+        public int Release1AddressLower => BitRange(863, 832);
+        public int Release1AddressUpper => BitRange(871, 864);
+        public int QmdReservedL => BitRange(879, 872);
+        public ReductionOp Release1ReductionOp => (ReductionOp)BitRange(886, 884);
+        public bool QmdReservedM => Bit(887);
+        public ReductionFormat Release1ReductionFormat => (ReductionFormat)BitRange(889, 888);
+        public bool Release1ReductionEnable => Bit(890);
+        public StructureSize Release1StructureSize => (StructureSize)BitRange(895, 895);
+        public int Release1Payload => BitRange(927, 896);
+        public int ConstantBufferAddrLower(int i) => BitRange(959 + i * 64, 928 + i * 64);
+        public int ConstantBufferAddrUpper(int i) => BitRange(967 + i * 64, 960 + i * 64);
+        public int ConstantBufferReservedAddr(int i) => BitRange(973 + i * 64, 968 + i * 64);
+        public bool ConstantBufferInvalidate(int i) => Bit(974 + i * 64);
+        public int ConstantBufferSize(int i) => BitRange(991 + i * 64, 975 + i * 64);
+        public int ShaderLocalMemoryLowSize => BitRange(1463, 1440);
+        public int QmdReservedN => BitRange(1466, 1464);
+        public int BarrierCount => BitRange(1471, 1467);
+        public int ShaderLocalMemoryHighSize => BitRange(1495, 1472);
+        public int RegisterCount => BitRange(1503, 1496);
+        public int ShaderLocalMemoryCrsSize => BitRange(1527, 1504);
+        public int SassVersion => BitRange(1535, 1528);
+        public int HwOnlyInnerGet => BitRange(1566, 1536);
+        public bool HwOnlyRequireSchedulingPcas => Bit(1567);
+        public int HwOnlyInnerPut => BitRange(1598, 1568);
+        public bool HwOnlyScgType => Bit(1599);
+        public int HwOnlySpanListHeadIndex => BitRange(1629, 1600);
+        public bool QmdReservedQ => Bit(1630);
+        public bool HwOnlySpanListHeadIndexValid => Bit(1631);
+        public int HwOnlySkedNextQmdPointer => BitRange(1663, 1632);
+        public int QmdSpareE => BitRange(1695, 1664);
+        public int QmdSpareF => BitRange(1727, 1696);
+        public int QmdSpareG => BitRange(1759, 1728);
+        public int QmdSpareH => BitRange(1791, 1760);
+        public int QmdSpareI => BitRange(1823, 1792);
+        public int QmdSpareJ => BitRange(1855, 1824);
+        public int QmdSpareK => BitRange(1887, 1856);
+        public int QmdSpareL => BitRange(1919, 1888);
+        public int QmdSpareM => BitRange(1951, 1920);
+        public int QmdSpareN => BitRange(1983, 1952);
+        public int DebugIdUpper => BitRange(2015, 1984);
+        public int DebugIdLower => BitRange(2047, 2016);
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private bool Bit(int bit)
+        {
+            if ((uint)bit >= 64 * 32)
+            {
+                throw new ArgumentOutOfRangeException(nameof(bit));
+            }
+
+            return (_words[bit >> 5] & (1 << (bit & 31))) != 0;
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private int BitRange(int upper, int lower)
+        {
+            if ((uint)lower >= 64 * 32)
+            {
+                throw new ArgumentOutOfRangeException(nameof(lower));
+            }
+
+            int mask = (int)(uint.MaxValue >> (32 - (upper - lower + 1)));
+
+            return (_words[lower >> 5] >> (lower & 31)) & mask;
+        }
+    }
+}
\ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/ConditionalRenderEnabled.cs b/src/Ryujinx.Graphics.Gpu/Engine/ConditionalRenderEnabled.cs
new file mode 100644
index 00000000..5581b5cc
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/ConditionalRenderEnabled.cs
@@ -0,0 +1,12 @@
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    /// <summary>
+    /// Conditional rendering enable.
+    /// </summary>
+    enum ConditionalRenderEnabled
+    {
+        False,
+        True,
+        Host
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/DeviceStateWithShadow.cs b/src/Ryujinx.Graphics.Gpu/Engine/DeviceStateWithShadow.cs
new file mode 100644
index 00000000..74a9aa04
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/DeviceStateWithShadow.cs
@@ -0,0 +1,96 @@
+using Ryujinx.Graphics.Device;
+using System;
+using System.Collections.Generic;
+using System.Diagnostics;
+using System.Diagnostics.CodeAnalysis;
+using System.Runtime.CompilerServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    /// <summary>
+    /// State interface with a shadow memory control register.
+    /// </summary>
+    interface IShadowState
+    {
+        /// <summary>
+        /// MME shadow ram control mode.
+        /// </summary>
+        SetMmeShadowRamControlMode SetMmeShadowRamControlMode { get; }
+    }
+
+    /// <summary>
+    /// Represents a device's state, with a additional shadow state.
+    /// </summary>
+    /// <typeparam name="TState">Type of the state</typeparam>
+    class DeviceStateWithShadow<[DynamicallyAccessedMembers(DynamicallyAccessedMemberTypes.PublicFields)] TState> : IDeviceState where TState : unmanaged, IShadowState
+    {
+        private readonly DeviceState<TState> _state;
+        private readonly DeviceState<TState> _shadowState;
+
+        /// <summary>
+        /// Current device state.
+        /// </summary>
+        public ref TState State => ref _state.State;
+
+        /// <summary>
+        /// Creates a new instance of the device state, with shadow state.
+        /// </summary>
+        /// <param name="callbacks">Optional that will be called if a register specified by name is read or written</param>
+        /// <param name="debugLogCallback">Optional callback to be used for debug log messages</param>
+        public DeviceStateWithShadow(IReadOnlyDictionary<string, RwCallback> callbacks = null, Action<string> debugLogCallback = null)
+        {
+            _state = new DeviceState<TState>(callbacks, debugLogCallback);
+            _shadowState = new DeviceState<TState>();
+        }
+
+        /// <summary>
+        /// Reads a value from a register.
+        /// </summary>
+        /// <param name="offset">Register offset in bytes</param>
+        /// <returns>Value stored on the register</returns>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public int Read(int offset)
+        {
+            return _state.Read(offset);
+        }
+
+        /// <summary>
+        /// Writes a value to a register.
+        /// </summary>
+        /// <param name="offset">Register offset in bytes</param>
+        /// <param name="value">Value to be written</param>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void Write(int offset, int value)
+        {
+            WriteWithRedundancyCheck(offset, value, out _);
+        }
+
+        /// <summary>
+        /// Writes a value to a register, returning a value indicating if <paramref name="value"/>
+        /// is different from the current value on the register.
+        /// </summary>
+        /// <param name="offset">Register offset in bytes</param>
+        /// <param name="value">Value to be written</param>
+        /// <param name="changed">True if the value was changed, false otherwise</param>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void WriteWithRedundancyCheck(int offset, int value, out bool changed)
+        {
+            var shadowRamControl = _state.State.SetMmeShadowRamControlMode;
+            if (shadowRamControl == SetMmeShadowRamControlMode.MethodPassthrough || offset < 0x200)
+            {
+                _state.WriteWithRedundancyCheck(offset, value, out changed);
+            }
+            else if (shadowRamControl == SetMmeShadowRamControlMode.MethodTrack ||
+                     shadowRamControl == SetMmeShadowRamControlMode.MethodTrackWithFilter)
+            {
+                _shadowState.Write(offset, value);
+                _state.WriteWithRedundancyCheck(offset, value, out changed);
+            }
+            else /* if (shadowRamControl == SetMmeShadowRamControlMode.MethodReplay) */
+            {
+                Debug.Assert(shadowRamControl == SetMmeShadowRamControlMode.MethodReplay);
+                _state.WriteWithRedundancyCheck(offset, _shadowState.Read(offset), out changed);
+            }
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Dma/DmaClass.cs b/src/Ryujinx.Graphics.Gpu/Engine/Dma/DmaClass.cs
new file mode 100644
index 00000000..fd93cd8b
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Dma/DmaClass.cs
@@ -0,0 +1,635 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.Device;
+using Ryujinx.Graphics.Gpu.Engine.Threed;
+using Ryujinx.Graphics.Gpu.Memory;
+using Ryujinx.Graphics.Texture;
+using System;
+using System.Collections.Generic;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+using System.Runtime.Intrinsics;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Dma
+{
+    /// <summary>
+    /// Represents a DMA copy engine class.
+    /// </summary>
+    class DmaClass : IDeviceState
+    {
+        private readonly GpuContext _context;
+        private readonly GpuChannel _channel;
+        private readonly ThreedClass _3dEngine;
+        private readonly DeviceState<DmaClassState> _state;
+
+        /// <summary>
+        /// Copy flags passed on DMA launch.
+        /// </summary>
+        [Flags]
+        private enum CopyFlags
+        {
+            SrcLinear = 1 << 7,
+            DstLinear = 1 << 8,
+            MultiLineEnable = 1 << 9,
+            RemapEnable = 1 << 10
+        }
+
+        /// <summary>
+        /// Texture parameters for copy.
+        /// </summary>
+        private struct TextureParams
+        {
+            /// <summary>
+            /// Copy region X coordinate.
+            /// </summary>
+            public readonly int RegionX;
+
+            /// <summary>
+            /// Copy region Y coordinate.
+            /// </summary>
+            public readonly int RegionY;
+
+            /// <summary>
+            /// Offset from the base pointer of the data in memory.
+            /// </summary>
+            public readonly int BaseOffset;
+
+            /// <summary>
+            /// Bytes per pixel.
+            /// </summary>
+            public readonly int Bpp;
+
+            /// <summary>
+            /// Whether the texture is linear. If false, the texture is block linear.
+            /// </summary>
+            public readonly bool Linear;
+
+            /// <summary>
+            /// Pixel offset from XYZ coordinates calculator.
+            /// </summary>
+            public readonly OffsetCalculator Calculator;
+
+            /// <summary>
+            /// Creates texture parameters.
+            /// </summary>
+            /// <param name="regionX">Copy region X coordinate</param>
+            /// <param name="regionY">Copy region Y coordinate</param>
+            /// <param name="baseOffset">Offset from the base pointer of the data in memory</param>
+            /// <param name="bpp">Bytes per pixel</param>
+            /// <param name="linear">Whether the texture is linear. If false, the texture is block linear</param>
+            /// <param name="calculator">Pixel offset from XYZ coordinates calculator</param>
+            public TextureParams(int regionX, int regionY, int baseOffset, int bpp, bool linear, OffsetCalculator calculator)
+            {
+                RegionX = regionX;
+                RegionY = regionY;
+                BaseOffset = baseOffset;
+                Bpp = bpp;
+                Linear = linear;
+                Calculator = calculator;
+            }
+        }
+
+        [StructLayout(LayoutKind.Sequential, Size = 3, Pack = 1)]
+        private struct UInt24
+        {
+            public byte Byte0;
+            public byte Byte1;
+            public byte Byte2;
+        }
+
+        /// <summary>
+        /// Creates a new instance of the DMA copy engine class.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        /// <param name="channel">GPU channel</param>
+        /// <param name="threedEngine">3D engine</param>
+        public DmaClass(GpuContext context, GpuChannel channel, ThreedClass threedEngine)
+        {
+            _context = context;
+            _channel = channel;
+            _3dEngine = threedEngine;
+            _state = new DeviceState<DmaClassState>(new Dictionary<string, RwCallback>
+            {
+                { nameof(DmaClassState.LaunchDma), new RwCallback(LaunchDma, null) }
+            });
+        }
+
+        /// <summary>
+        /// Reads data from the class registers.
+        /// </summary>
+        /// <param name="offset">Register byte offset</param>
+        /// <returns>Data at the specified offset</returns>
+        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>
+        public void Write(int offset, int data) => _state.Write(offset, data);
+
+        /// <summary>
+        /// Determine if a buffer-to-texture region covers the entirety of a texture.
+        /// </summary>
+        /// <param name="tex">Texture to compare</param>
+        /// <param name="linear">True if the texture is linear, false if block linear</param>
+        /// <param name="bpp">Texture bytes per pixel</param>
+        /// <param name="stride">Texture stride</param>
+        /// <param name="xCount">Number of pixels to be copied</param>
+        /// <param name="yCount">Number of lines to be copied</param>
+        /// <returns></returns>
+        private static bool IsTextureCopyComplete(DmaTexture tex, bool linear, int bpp, int stride, int xCount, int yCount)
+        {
+            if (linear)
+            {
+                // If the stride is negative, the texture has to be flipped, so
+                // the fast copy is not trivial, use the slow path.
+                if (stride <= 0)
+                {
+                    return false;
+                }
+
+                int alignWidth = Constants.StrideAlignment / bpp;
+                return stride / bpp == BitUtils.AlignUp(xCount, alignWidth);
+            }
+            else
+            {
+                int alignWidth = Constants.GobAlignment / bpp;
+                return tex.RegionX == 0 &&
+                       tex.RegionY == 0 &&
+                       tex.Width == BitUtils.AlignUp(xCount, alignWidth) &&
+                       tex.Height == yCount;
+            }
+        }
+
+        /// <summary>
+        /// Releases a semaphore for a given LaunchDma method call.
+        /// </summary>
+        /// <param name="argument">The LaunchDma call argument</param>
+        private void ReleaseSemaphore(int argument)
+        {
+            LaunchDmaSemaphoreType type = (LaunchDmaSemaphoreType)((argument >> 3) & 0x3);
+            if (type != LaunchDmaSemaphoreType.None)
+            {
+                ulong address = ((ulong)_state.State.SetSemaphoreA << 32) | _state.State.SetSemaphoreB;
+                if (type == LaunchDmaSemaphoreType.ReleaseOneWordSemaphore)
+                {
+                    _channel.MemoryManager.Write(address, _state.State.SetSemaphorePayload);
+                }
+                else /* if (type == LaunchDmaSemaphoreType.ReleaseFourWordSemaphore) */
+                {
+                    _channel.MemoryManager.Write(address + 8, _context.GetTimestamp());
+                    _channel.MemoryManager.Write(address, (ulong)_state.State.SetSemaphorePayload);
+                }
+            }
+        }
+
+        /// <summary>
+        /// Performs a buffer to buffer, or buffer to texture copy.
+        /// </summary>
+        /// <param name="argument">The LaunchDma call argument</param>
+        private void DmaCopy(int argument)
+        {
+            var memoryManager = _channel.MemoryManager;
+
+            CopyFlags copyFlags = (CopyFlags)argument;
+
+            bool srcLinear = copyFlags.HasFlag(CopyFlags.SrcLinear);
+            bool dstLinear = copyFlags.HasFlag(CopyFlags.DstLinear);
+            bool copy2D = copyFlags.HasFlag(CopyFlags.MultiLineEnable);
+            bool remap = copyFlags.HasFlag(CopyFlags.RemapEnable);
+
+            uint size = _state.State.LineLengthIn;
+
+            if (size == 0)
+            {
+                return;
+            }
+
+            ulong srcGpuVa = ((ulong)_state.State.OffsetInUpperUpper << 32) | _state.State.OffsetInLower;
+            ulong dstGpuVa = ((ulong)_state.State.OffsetOutUpperUpper << 32) | _state.State.OffsetOutLower;
+
+            int xCount = (int)_state.State.LineLengthIn;
+            int yCount = (int)_state.State.LineCount;
+
+            _3dEngine.CreatePendingSyncs();
+            _3dEngine.FlushUboDirty();
+
+            if (copy2D)
+            {
+                // Buffer to texture copy.
+                int componentSize = (int)_state.State.SetRemapComponentsComponentSize + 1;
+                int srcComponents = (int)_state.State.SetRemapComponentsNumSrcComponents + 1;
+                int dstComponents = (int)_state.State.SetRemapComponentsNumDstComponents + 1;
+                int srcBpp = remap ? srcComponents * componentSize : 1;
+                int dstBpp = remap ? dstComponents * componentSize : 1;
+
+                var dst = Unsafe.As<uint, DmaTexture>(ref _state.State.SetDstBlockSize);
+                var src = Unsafe.As<uint, DmaTexture>(ref _state.State.SetSrcBlockSize);
+
+                int srcRegionX = 0, srcRegionY = 0, dstRegionX = 0, dstRegionY = 0;
+
+                if (!srcLinear)
+                {
+                    srcRegionX = src.RegionX;
+                    srcRegionY = src.RegionY;
+                }
+
+                if (!dstLinear)
+                {
+                    dstRegionX = dst.RegionX;
+                    dstRegionY = dst.RegionY;
+                }
+
+                int srcStride = (int)_state.State.PitchIn;
+                int dstStride = (int)_state.State.PitchOut;
+
+                var srcCalculator = new OffsetCalculator(
+                    src.Width,
+                    src.Height,
+                    srcStride,
+                    srcLinear,
+                    src.MemoryLayout.UnpackGobBlocksInY(),
+                    src.MemoryLayout.UnpackGobBlocksInZ(),
+                    srcBpp);
+
+                var dstCalculator = new OffsetCalculator(
+                    dst.Width,
+                    dst.Height,
+                    dstStride,
+                    dstLinear,
+                    dst.MemoryLayout.UnpackGobBlocksInY(),
+                    dst.MemoryLayout.UnpackGobBlocksInZ(),
+                    dstBpp);
+
+                (int srcBaseOffset, int srcSize) = srcCalculator.GetRectangleRange(srcRegionX, srcRegionY, xCount, yCount);
+                (int dstBaseOffset, int dstSize) = dstCalculator.GetRectangleRange(dstRegionX, dstRegionY, xCount, yCount);
+
+                if (srcLinear && srcStride < 0)
+                {
+                    srcBaseOffset += srcStride * (yCount - 1);
+                }
+
+                if (dstLinear && dstStride < 0)
+                {
+                    dstBaseOffset += dstStride * (yCount - 1);
+                }
+
+                ReadOnlySpan<byte> srcSpan = memoryManager.GetSpan(srcGpuVa + (ulong)srcBaseOffset, srcSize, true);
+
+                bool completeSource = IsTextureCopyComplete(src, srcLinear, srcBpp, srcStride, xCount, yCount);
+                bool completeDest = IsTextureCopyComplete(dst, dstLinear, dstBpp, dstStride, xCount, yCount);
+
+                if (completeSource && completeDest)
+                {
+                    var target = memoryManager.Physical.TextureCache.FindTexture(
+                        memoryManager,
+                        dstGpuVa,
+                        dstBpp,
+                        dstStride,
+                        dst.Height,
+                        xCount,
+                        yCount,
+                        dstLinear,
+                        dst.MemoryLayout.UnpackGobBlocksInY(),
+                        dst.MemoryLayout.UnpackGobBlocksInZ());
+
+                    if (target != null)
+                    {
+                        byte[] data;
+                        if (srcLinear)
+                        {
+                            data = LayoutConverter.ConvertLinearStridedToLinear(
+                                target.Info.Width,
+                                target.Info.Height,
+                                1,
+                                1,
+                                xCount * srcBpp,
+                                srcStride,
+                                target.Info.FormatInfo.BytesPerPixel,
+                                srcSpan);
+                        }
+                        else
+                        {
+                            data = LayoutConverter.ConvertBlockLinearToLinear(
+                                src.Width,
+                                src.Height,
+                                src.Depth,
+                                1,
+                                1,
+                                1,
+                                1,
+                                1,
+                                srcBpp,
+                                src.MemoryLayout.UnpackGobBlocksInY(),
+                                src.MemoryLayout.UnpackGobBlocksInZ(),
+                                1,
+                                new SizeInfo((int)target.Size),
+                                srcSpan);
+                        }
+
+                        target.SynchronizeMemory();
+                        target.SetData(data);
+                        target.SignalModified();
+                        return;
+                    }
+                    else if (srcCalculator.LayoutMatches(dstCalculator))
+                    {
+                        // No layout conversion has to be performed, just copy the data entirely.
+                        memoryManager.Write(dstGpuVa + (ulong)dstBaseOffset, srcSpan);
+                        return;
+                    }
+                }
+
+                // OPT: This allocates a (potentially) huge temporary array and then copies an existing
+                // region of memory into it, data that might get overwritten entirely anyways. Ideally this should
+                // all be rewritten to use pooled arrays, but that gets complicated with packed data and strides
+                Span<byte> dstSpan = memoryManager.GetSpan(dstGpuVa + (ulong)dstBaseOffset, dstSize).ToArray();
+
+                TextureParams srcParams = new TextureParams(srcRegionX, srcRegionY, srcBaseOffset, srcBpp, srcLinear, srcCalculator);
+                TextureParams dstParams = new TextureParams(dstRegionX, dstRegionY, dstBaseOffset, dstBpp, dstLinear, dstCalculator);
+
+                // If remapping is enabled, we always copy the components directly, in order.
+                // If it's enabled, but the mapping is just XYZW, we also copy them in order.
+                bool isIdentityRemap = !remap ||
+                    (_state.State.SetRemapComponentsDstX == SetRemapComponentsDst.SrcX &&
+                    (dstComponents < 2 || _state.State.SetRemapComponentsDstY == SetRemapComponentsDst.SrcY) &&
+                    (dstComponents < 3 || _state.State.SetRemapComponentsDstZ == SetRemapComponentsDst.SrcZ) &&
+                    (dstComponents < 4 || _state.State.SetRemapComponentsDstW == SetRemapComponentsDst.SrcW));
+
+                if (isIdentityRemap)
+                {
+                    // The order of the components doesn't change, so we can just copy directly
+                    // (with layout conversion if necessary).
+
+                    switch (srcBpp)
+                    {
+                        case 1: Copy<byte>(dstSpan, srcSpan, dstParams, srcParams); break;
+                        case 2: Copy<ushort>(dstSpan, srcSpan, dstParams, srcParams); break;
+                        case 4: Copy<uint>(dstSpan, srcSpan, dstParams, srcParams); break;
+                        case 8: Copy<ulong>(dstSpan, srcSpan, dstParams, srcParams); break;
+                        case 12: Copy<Bpp12Pixel>(dstSpan, srcSpan, dstParams, srcParams); break;
+                        case 16: Copy<Vector128<byte>>(dstSpan, srcSpan, dstParams, srcParams); break;
+                        default: throw new NotSupportedException($"Unable to copy ${srcBpp} bpp pixel format.");
+                    }
+                }
+                else
+                {
+                    // The order or value of the components might change.
+
+                    switch (componentSize)
+                    {
+                        case 1: CopyShuffle<byte>(dstSpan, srcSpan, dstParams, srcParams); break;
+                        case 2: CopyShuffle<ushort>(dstSpan, srcSpan, dstParams, srcParams); break;
+                        case 3: CopyShuffle<UInt24>(dstSpan, srcSpan, dstParams, srcParams); break;
+                        case 4: CopyShuffle<uint>(dstSpan, srcSpan, dstParams, srcParams); break;
+                        default: throw new NotSupportedException($"Unable to copy ${componentSize} component size.");
+                    }
+                }
+
+                memoryManager.Write(dstGpuVa + (ulong)dstBaseOffset, dstSpan);
+            }
+            else
+            {
+                if (remap &&
+                    _state.State.SetRemapComponentsDstX == SetRemapComponentsDst.ConstA &&
+                    _state.State.SetRemapComponentsDstY == SetRemapComponentsDst.ConstA &&
+                    _state.State.SetRemapComponentsDstZ == SetRemapComponentsDst.ConstA &&
+                    _state.State.SetRemapComponentsDstW == SetRemapComponentsDst.ConstA &&
+                    _state.State.SetRemapComponentsNumSrcComponents == SetRemapComponentsNumComponents.One &&
+                    _state.State.SetRemapComponentsNumDstComponents == SetRemapComponentsNumComponents.One &&
+                    _state.State.SetRemapComponentsComponentSize == SetRemapComponentsComponentSize.Four)
+                {
+                    // Fast path for clears when remap is enabled.
+                    memoryManager.Physical.BufferCache.ClearBuffer(memoryManager, dstGpuVa, size * 4, _state.State.SetRemapConstA);
+                }
+                else
+                {
+                    // TODO: Implement remap functionality.
+                    // Buffer to buffer copy.
+
+                    bool srcIsPitchKind = memoryManager.GetKind(srcGpuVa).IsPitch();
+                    bool dstIsPitchKind = memoryManager.GetKind(dstGpuVa).IsPitch();
+
+                    if (!srcIsPitchKind && dstIsPitchKind)
+                    {
+                        CopyGobBlockLinearToLinear(memoryManager, srcGpuVa, dstGpuVa, size);
+                    }
+                    else if (srcIsPitchKind && !dstIsPitchKind)
+                    {
+                        CopyGobLinearToBlockLinear(memoryManager, srcGpuVa, dstGpuVa, size);
+                    }
+                    else
+                    {
+                        memoryManager.Physical.BufferCache.CopyBuffer(memoryManager, srcGpuVa, dstGpuVa, size);
+                    }
+                }
+            }
+        }
+
+        /// <summary>
+        /// Copies data from one texture to another, while performing layout conversion if necessary.
+        /// </summary>
+        /// <typeparam name="T">Pixel type</typeparam>
+        /// <param name="dstSpan">Destination texture memory region</param>
+        /// <param name="srcSpan">Source texture memory region</param>
+        /// <param name="dst">Destination texture parameters</param>
+        /// <param name="src">Source texture parameters</param>
+        private unsafe void Copy<T>(Span<byte> dstSpan, ReadOnlySpan<byte> srcSpan, TextureParams dst, TextureParams src) where T : unmanaged
+        {
+            int xCount = (int)_state.State.LineLengthIn;
+            int yCount = (int)_state.State.LineCount;
+
+            if (src.Linear && dst.Linear && src.Bpp == dst.Bpp)
+            {
+                // Optimized path for purely linear copies - we don't need to calculate every single byte offset,
+                // and we can make use of Span.CopyTo which is very very fast (even compared to pointers)
+                for (int y = 0; y < yCount; y++)
+                {
+                    src.Calculator.SetY(src.RegionY + y);
+                    dst.Calculator.SetY(dst.RegionY + y);
+                    int srcOffset = src.Calculator.GetOffset(src.RegionX);
+                    int dstOffset = dst.Calculator.GetOffset(dst.RegionX);
+                    srcSpan.Slice(srcOffset - src.BaseOffset, xCount * src.Bpp)
+                        .CopyTo(dstSpan.Slice(dstOffset - dst.BaseOffset, xCount * dst.Bpp));
+                }
+            }
+            else
+            {
+                fixed (byte* dstPtr = dstSpan, srcPtr = srcSpan)
+                {
+                    byte* dstBase = dstPtr - dst.BaseOffset; // Layout offset is relative to the base, so we need to subtract the span's offset.
+                    byte* srcBase = srcPtr - src.BaseOffset;
+
+                    for (int y = 0; y < yCount; y++)
+                    {
+                        src.Calculator.SetY(src.RegionY + y);
+                        dst.Calculator.SetY(dst.RegionY + y);
+
+                        for (int x = 0; x < xCount; x++)
+                        {
+                            int srcOffset = src.Calculator.GetOffset(src.RegionX + x);
+                            int dstOffset = dst.Calculator.GetOffset(dst.RegionX + x);
+
+                            *(T*)(dstBase + dstOffset) = *(T*)(srcBase + srcOffset);
+                        }
+                    }
+                }
+            }
+        }
+
+        /// <summary>
+        /// Sets texture pixel data to a constant value, while performing layout conversion if necessary.
+        /// </summary>
+        /// <typeparam name="T">Pixel type</typeparam>
+        /// <param name="dstSpan">Destination texture memory region</param>
+        /// <param name="dst">Destination texture parameters</param>
+        /// <param name="fillValue">Constant pixel value to be set</param>
+        private unsafe void Fill<T>(Span<byte> dstSpan, TextureParams dst, T fillValue) where T : unmanaged
+        {
+            int xCount = (int)_state.State.LineLengthIn;
+            int yCount = (int)_state.State.LineCount;
+
+            fixed (byte* dstPtr = dstSpan)
+            {
+                byte* dstBase = dstPtr - dst.BaseOffset; // Layout offset is relative to the base, so we need to subtract the span's offset.
+
+                for (int y = 0; y < yCount; y++)
+                {
+                    dst.Calculator.SetY(dst.RegionY + y);
+
+                    for (int x = 0; x < xCount; x++)
+                    {
+                        int dstOffset = dst.Calculator.GetOffset(dst.RegionX + x);
+
+                        *(T*)(dstBase + dstOffset) = fillValue;
+                    }
+                }
+            }
+        }
+
+        /// <summary>
+        /// Copies data from one texture to another, while performing layout conversion and component shuffling if necessary.
+        /// </summary>
+        /// <typeparam name="T">Pixel type</typeparam>
+        /// <param name="dstSpan">Destination texture memory region</param>
+        /// <param name="srcSpan">Source texture memory region</param>
+        /// <param name="dst">Destination texture parameters</param>
+        /// <param name="src">Source texture parameters</param>
+        private void CopyShuffle<T>(Span<byte> dstSpan, ReadOnlySpan<byte> srcSpan, TextureParams dst, TextureParams src) where T : unmanaged
+        {
+            int dstComponents = (int)_state.State.SetRemapComponentsNumDstComponents + 1;
+
+            for (int i = 0; i < dstComponents; i++)
+            {
+                SetRemapComponentsDst componentsDst = i switch
+                {
+                    0 => _state.State.SetRemapComponentsDstX,
+                    1 => _state.State.SetRemapComponentsDstY,
+                    2 => _state.State.SetRemapComponentsDstZ,
+                    _ => _state.State.SetRemapComponentsDstW
+                };
+
+                switch (componentsDst)
+                {
+                    case SetRemapComponentsDst.SrcX:
+                        Copy<T>(dstSpan.Slice(Unsafe.SizeOf<T>() * i), srcSpan, dst, src);
+                        break;
+                    case SetRemapComponentsDst.SrcY:
+                        Copy<T>(dstSpan.Slice(Unsafe.SizeOf<T>() * i), srcSpan.Slice(Unsafe.SizeOf<T>()), dst, src);
+                        break;
+                    case SetRemapComponentsDst.SrcZ:
+                        Copy<T>(dstSpan.Slice(Unsafe.SizeOf<T>() * i), srcSpan.Slice(Unsafe.SizeOf<T>() * 2), dst, src);
+                        break;
+                    case SetRemapComponentsDst.SrcW:
+                        Copy<T>(dstSpan.Slice(Unsafe.SizeOf<T>() * i), srcSpan.Slice(Unsafe.SizeOf<T>() * 3), dst, src);
+                        break;
+                    case SetRemapComponentsDst.ConstA:
+                        Fill<T>(dstSpan.Slice(Unsafe.SizeOf<T>() * i), dst, Unsafe.As<uint, T>(ref _state.State.SetRemapConstA));
+                        break;
+                    case SetRemapComponentsDst.ConstB:
+                        Fill<T>(dstSpan.Slice(Unsafe.SizeOf<T>() * i), dst, Unsafe.As<uint, T>(ref _state.State.SetRemapConstB));
+                        break;
+                }
+            }
+        }
+
+        /// <summary>
+        /// Copies block linear data with block linear GOBs to a block linear destination with linear GOBs.
+        /// </summary>
+        /// <param name="memoryManager">GPU memory manager</param>
+        /// <param name="srcGpuVa">Source GPU virtual address</param>
+        /// <param name="dstGpuVa">Destination GPU virtual address</param>
+        /// <param name="size">Size in bytes of the copy</param>
+        private static void CopyGobBlockLinearToLinear(MemoryManager memoryManager, ulong srcGpuVa, ulong dstGpuVa, ulong size)
+        {
+            if (((srcGpuVa | dstGpuVa | size) & 0xf) == 0)
+            {
+                for (ulong offset = 0; offset < size; offset += 16)
+                {
+                    Vector128<byte> data = memoryManager.Read<Vector128<byte>>(ConvertGobLinearToBlockLinearAddress(srcGpuVa + offset), true);
+                    memoryManager.Write(dstGpuVa + offset, data);
+                }
+            }
+            else
+            {
+                for (ulong offset = 0; offset < size; offset++)
+                {
+                    byte data = memoryManager.Read<byte>(ConvertGobLinearToBlockLinearAddress(srcGpuVa + offset), true);
+                    memoryManager.Write(dstGpuVa + offset, data);
+                }
+            }
+        }
+
+        /// <summary>
+        /// Copies block linear data with linear GOBs to a block linear destination with block linear GOBs.
+        /// </summary>
+        /// <param name="memoryManager">GPU memory manager</param>
+        /// <param name="srcGpuVa">Source GPU virtual address</param>
+        /// <param name="dstGpuVa">Destination GPU virtual address</param>
+        /// <param name="size">Size in bytes of the copy</param>
+        private static void CopyGobLinearToBlockLinear(MemoryManager memoryManager, ulong srcGpuVa, ulong dstGpuVa, ulong size)
+        {
+            if (((srcGpuVa | dstGpuVa | size) & 0xf) == 0)
+            {
+                for (ulong offset = 0; offset < size; offset += 16)
+                {
+                    Vector128<byte> data = memoryManager.Read<Vector128<byte>>(srcGpuVa + offset, true);
+                    memoryManager.Write(ConvertGobLinearToBlockLinearAddress(dstGpuVa + offset), data);
+                }
+            }
+            else
+            {
+                for (ulong offset = 0; offset < size; offset++)
+                {
+                    byte data = memoryManager.Read<byte>(srcGpuVa + offset, true);
+                    memoryManager.Write(ConvertGobLinearToBlockLinearAddress(dstGpuVa + offset), data);
+                }
+            }
+        }
+
+        /// <summary>
+        /// Calculates the GOB block linear address from a linear address.
+        /// </summary>
+        /// <param name="address">Linear address</param>
+        /// <returns>Block linear address</returns>
+        private static ulong ConvertGobLinearToBlockLinearAddress(ulong address)
+        {
+            // y2 y1 y0 x5 x4 x3 x2 x1 x0 -> x5 y2 y1 x4 y0 x3 x2 x1 x0
+            return (address & ~0x1f0UL) |
+                ((address & 0x40) >> 2) |
+                ((address & 0x10) << 1) |
+                ((address & 0x180) >> 1) |
+                ((address & 0x20) << 3);
+        }
+
+        /// <summary>
+        /// Performs a buffer to buffer, or buffer to texture copy, then optionally releases a semaphore.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void LaunchDma(int argument)
+        {
+            DmaCopy(argument);
+            ReleaseSemaphore(argument);
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Dma/DmaClassState.cs b/src/Ryujinx.Graphics.Gpu/Engine/Dma/DmaClassState.cs
new file mode 100644
index 00000000..7de4d5f0
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Dma/DmaClassState.cs
@@ -0,0 +1,271 @@
+// This file was auto-generated from NVIDIA official Maxwell definitions.
+
+namespace Ryujinx.Graphics.Gpu.Engine.Dma
+{
+    /// <summary>
+    /// Physical mode target.
+    /// </summary>
+    enum SetPhysModeTarget
+    {
+        LocalFb = 0,
+        CoherentSysmem = 1,
+        NoncoherentSysmem = 2,
+    }
+
+    /// <summary>
+    /// DMA data transfer type.
+    /// </summary>
+    enum LaunchDmaDataTransferType
+    {
+        None = 0,
+        Pipelined = 1,
+        NonPipelined = 2,
+    }
+
+    /// <summary>
+    /// DMA semaphore type.
+    /// </summary>
+    enum LaunchDmaSemaphoreType
+    {
+        None = 0,
+        ReleaseOneWordSemaphore = 1,
+        ReleaseFourWordSemaphore = 2,
+    }
+
+    /// <summary>
+    /// DMA interrupt type.
+    /// </summary>
+    enum LaunchDmaInterruptType
+    {
+        None = 0,
+        Blocking = 1,
+        NonBlocking = 2,
+    }
+
+    /// <summary>
+    /// DMA destination memory layout.
+    /// </summary>
+    enum LaunchDmaMemoryLayout
+    {
+        Blocklinear = 0,
+        Pitch = 1,
+    }
+
+    /// <summary>
+    /// DMA type.
+    /// </summary>
+    enum LaunchDmaType
+    {
+        Virtual = 0,
+        Physical = 1,
+    }
+
+    /// <summary>
+    /// DMA semaphore reduction operation.
+    /// </summary>
+    enum LaunchDmaSemaphoreReduction
+    {
+        Imin = 0,
+        Imax = 1,
+        Ixor = 2,
+        Iand = 3,
+        Ior = 4,
+        Iadd = 5,
+        Inc = 6,
+        Dec = 7,
+        Fadd = 10,
+    }
+
+    /// <summary>
+    /// DMA semaphore reduction signedness.
+    /// </summary>
+    enum LaunchDmaSemaphoreReductionSign
+    {
+        Signed = 0,
+        Unsigned = 1,
+    }
+
+    /// <summary>
+    /// DMA L2 cache bypass.
+    /// </summary>
+    enum LaunchDmaBypassL2
+    {
+        UsePteSetting = 0,
+        ForceVolatile = 1,
+    }
+
+    /// <summary>
+    /// DMA component remapping source component.
+    /// </summary>
+    enum SetRemapComponentsDst
+    {
+        SrcX = 0,
+        SrcY = 1,
+        SrcZ = 2,
+        SrcW = 3,
+        ConstA = 4,
+        ConstB = 5,
+        NoWrite = 6,
+    }
+
+    /// <summary>
+    /// DMA component remapping component size.
+    /// </summary>
+    enum SetRemapComponentsComponentSize
+    {
+        One = 0,
+        Two = 1,
+        Three = 2,
+        Four = 3,
+    }
+
+    /// <summary>
+    /// DMA component remapping number of components.
+    /// </summary>
+    enum SetRemapComponentsNumComponents
+    {
+        One = 0,
+        Two = 1,
+        Three = 2,
+        Four = 3,
+    }
+
+    /// <summary>
+    /// Width in GOBs of the destination texture.
+    /// </summary>
+    enum SetBlockSizeWidth
+    {
+        QuarterGob = 14,
+        OneGob = 0,
+    }
+
+    /// <summary>
+    /// Height in GOBs of the destination texture.
+    /// </summary>
+    enum SetBlockSizeHeight
+    {
+        OneGob = 0,
+        TwoGobs = 1,
+        FourGobs = 2,
+        EightGobs = 3,
+        SixteenGobs = 4,
+        ThirtytwoGobs = 5,
+    }
+
+    /// <summary>
+    /// Depth in GOBs of the destination texture.
+    /// </summary>
+    enum SetBlockSizeDepth
+    {
+        OneGob = 0,
+        TwoGobs = 1,
+        FourGobs = 2,
+        EightGobs = 3,
+        SixteenGobs = 4,
+        ThirtytwoGobs = 5,
+    }
+
+    /// <summary>
+    /// Height of a single GOB in lines.
+    /// </summary>
+    enum SetBlockSizeGobHeight
+    {
+        GobHeightTesla4 = 0,
+        GobHeightFermi8 = 1,
+    }
+
+    /// <summary>
+    /// DMA copy class state.
+    /// </summary>
+    unsafe struct DmaClassState
+    {
+#pragma warning disable CS0649
+        public fixed uint Reserved00[64];
+        public uint Nop;
+        public fixed uint Reserved104[15];
+        public uint PmTrigger;
+        public fixed uint Reserved144[63];
+        public uint SetSemaphoreA;
+        public int SetSemaphoreAUpper => (int)((SetSemaphoreA >> 0) & 0xFF);
+        public uint SetSemaphoreB;
+        public uint SetSemaphorePayload;
+        public fixed uint Reserved24C[2];
+        public uint SetRenderEnableA;
+        public int SetRenderEnableAUpper => (int)((SetRenderEnableA >> 0) & 0xFF);
+        public uint SetRenderEnableB;
+        public uint SetRenderEnableC;
+        public int SetRenderEnableCMode => (int)((SetRenderEnableC >> 0) & 0x7);
+        public uint SetSrcPhysMode;
+        public SetPhysModeTarget SetSrcPhysModeTarget => (SetPhysModeTarget)((SetSrcPhysMode >> 0) & 0x3);
+        public uint SetDstPhysMode;
+        public SetPhysModeTarget SetDstPhysModeTarget => (SetPhysModeTarget)((SetDstPhysMode >> 0) & 0x3);
+        public fixed uint Reserved268[38];
+        public uint LaunchDma;
+        public LaunchDmaDataTransferType LaunchDmaDataTransferType => (LaunchDmaDataTransferType)((LaunchDma >> 0) & 0x3);
+        public bool LaunchDmaFlushEnable => (LaunchDma & 0x4) != 0;
+        public LaunchDmaSemaphoreType LaunchDmaSemaphoreType => (LaunchDmaSemaphoreType)((LaunchDma >> 3) & 0x3);
+        public LaunchDmaInterruptType LaunchDmaInterruptType => (LaunchDmaInterruptType)((LaunchDma >> 5) & 0x3);
+        public LaunchDmaMemoryLayout LaunchDmaSrcMemoryLayout => (LaunchDmaMemoryLayout)((LaunchDma >> 7) & 0x1);
+        public LaunchDmaMemoryLayout LaunchDmaDstMemoryLayout => (LaunchDmaMemoryLayout)((LaunchDma >> 8) & 0x1);
+        public bool LaunchDmaMultiLineEnable => (LaunchDma & 0x200) != 0;
+        public bool LaunchDmaRemapEnable => (LaunchDma & 0x400) != 0;
+        public bool LaunchDmaForceRmwdisable => (LaunchDma & 0x800) != 0;
+        public LaunchDmaType LaunchDmaSrcType => (LaunchDmaType)((LaunchDma >> 12) & 0x1);
+        public LaunchDmaType LaunchDmaDstType => (LaunchDmaType)((LaunchDma >> 13) & 0x1);
+        public LaunchDmaSemaphoreReduction LaunchDmaSemaphoreReduction => (LaunchDmaSemaphoreReduction)((LaunchDma >> 14) & 0xF);
+        public LaunchDmaSemaphoreReductionSign LaunchDmaSemaphoreReductionSign => (LaunchDmaSemaphoreReductionSign)((LaunchDma >> 18) & 0x1);
+        public bool LaunchDmaSemaphoreReductionEnable => (LaunchDma & 0x80000) != 0;
+        public LaunchDmaBypassL2 LaunchDmaBypassL2 => (LaunchDmaBypassL2)((LaunchDma >> 20) & 0x1);
+        public fixed uint Reserved304[63];
+        public uint OffsetInUpper;
+        public int OffsetInUpperUpper => (int)((OffsetInUpper >> 0) & 0xFF);
+        public uint OffsetInLower;
+        public uint OffsetOutUpper;
+        public int OffsetOutUpperUpper => (int)((OffsetOutUpper >> 0) & 0xFF);
+        public uint OffsetOutLower;
+        public uint PitchIn;
+        public uint PitchOut;
+        public uint LineLengthIn;
+        public uint LineCount;
+        public fixed uint Reserved420[184];
+        public uint SetRemapConstA;
+        public uint SetRemapConstB;
+        public uint SetRemapComponents;
+        public SetRemapComponentsDst SetRemapComponentsDstX => (SetRemapComponentsDst)((SetRemapComponents >> 0) & 0x7);
+        public SetRemapComponentsDst SetRemapComponentsDstY => (SetRemapComponentsDst)((SetRemapComponents >> 4) & 0x7);
+        public SetRemapComponentsDst SetRemapComponentsDstZ => (SetRemapComponentsDst)((SetRemapComponents >> 8) & 0x7);
+        public SetRemapComponentsDst SetRemapComponentsDstW => (SetRemapComponentsDst)((SetRemapComponents >> 12) & 0x7);
+        public SetRemapComponentsComponentSize SetRemapComponentsComponentSize => (SetRemapComponentsComponentSize)((SetRemapComponents >> 16) & 0x3);
+        public SetRemapComponentsNumComponents SetRemapComponentsNumSrcComponents => (SetRemapComponentsNumComponents)((SetRemapComponents >> 20) & 0x3);
+        public SetRemapComponentsNumComponents SetRemapComponentsNumDstComponents => (SetRemapComponentsNumComponents)((SetRemapComponents >> 24) & 0x3);
+        public uint SetDstBlockSize;
+        public SetBlockSizeWidth SetDstBlockSizeWidth => (SetBlockSizeWidth)((SetDstBlockSize >> 0) & 0xF);
+        public SetBlockSizeHeight SetDstBlockSizeHeight => (SetBlockSizeHeight)((SetDstBlockSize >> 4) & 0xF);
+        public SetBlockSizeDepth SetDstBlockSizeDepth => (SetBlockSizeDepth)((SetDstBlockSize >> 8) & 0xF);
+        public SetBlockSizeGobHeight SetDstBlockSizeGobHeight => (SetBlockSizeGobHeight)((SetDstBlockSize >> 12) & 0xF);
+        public uint SetDstWidth;
+        public uint SetDstHeight;
+        public uint SetDstDepth;
+        public uint SetDstLayer;
+        public uint SetDstOrigin;
+        public int SetDstOriginX => (int)((SetDstOrigin >> 0) & 0xFFFF);
+        public int SetDstOriginY => (int)((SetDstOrigin >> 16) & 0xFFFF);
+        public uint Reserved724;
+        public uint SetSrcBlockSize;
+        public SetBlockSizeWidth SetSrcBlockSizeWidth => (SetBlockSizeWidth)((SetSrcBlockSize >> 0) & 0xF);
+        public SetBlockSizeHeight SetSrcBlockSizeHeight => (SetBlockSizeHeight)((SetSrcBlockSize >> 4) & 0xF);
+        public SetBlockSizeDepth SetSrcBlockSizeDepth => (SetBlockSizeDepth)((SetSrcBlockSize >> 8) & 0xF);
+        public SetBlockSizeGobHeight SetSrcBlockSizeGobHeight => (SetBlockSizeGobHeight)((SetSrcBlockSize >> 12) & 0xF);
+        public uint SetSrcWidth;
+        public uint SetSrcHeight;
+        public uint SetSrcDepth;
+        public uint SetSrcLayer;
+        public uint SetSrcOrigin;
+        public int SetSrcOriginX => (int)((SetSrcOrigin >> 0) & 0xFFFF);
+        public int SetSrcOriginY => (int)((SetSrcOrigin >> 16) & 0xFFFF);
+        public fixed uint Reserved740[629];
+        public uint PmTriggerEnd;
+        public fixed uint Reserved1118[2490];
+#pragma warning restore CS0649
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Dma/DmaTexture.cs b/src/Ryujinx.Graphics.Gpu/Engine/Dma/DmaTexture.cs
new file mode 100644
index 00000000..6873ff40
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Dma/DmaTexture.cs
@@ -0,0 +1,20 @@
+using Ryujinx.Graphics.Gpu.Engine.Types;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Dma
+{
+    /// <summary>
+    /// Buffer to texture copy parameters.
+    /// </summary>
+    struct DmaTexture
+    {
+#pragma warning disable CS0649
+        public MemoryLayout MemoryLayout;
+        public int Width;
+        public int Height;
+        public int Depth;
+        public int RegionZ;
+        public ushort RegionX;
+        public ushort RegionY;
+#pragma warning restore CS0649
+    }
+}
\ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/CompressedMethod.cs b/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/CompressedMethod.cs
new file mode 100644
index 00000000..458dc8f6
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/CompressedMethod.cs
@@ -0,0 +1,41 @@
+// This file was auto-generated from NVIDIA official Maxwell definitions.
+
+namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
+{
+    enum TertOp
+    {
+        Grp0IncMethod = 0,
+        Grp0SetSubDevMask = 1,
+        Grp0StoreSubDevMask = 2,
+        Grp0UseSubDevMask = 3,
+        Grp2NonIncMethod = 0
+    }
+
+    enum SecOp
+    {
+        Grp0UseTert = 0,
+        IncMethod = 1,
+        Grp2UseTert = 2,
+        NonIncMethod = 3,
+        ImmdDataMethod = 4,
+        OneInc = 5,
+        Reserved6 = 6,
+        EndPbSegment = 7
+    }
+
+    struct CompressedMethod
+    {
+#pragma warning disable CS0649
+        public uint Method;
+#pragma warning restore CS0649
+        public int MethodAddressOld => (int)((Method >> 2) & 0x7FF);
+        public int MethodAddress => (int)((Method >> 0) & 0xFFF);
+        public int SubdeviceMask => (int)((Method >> 4) & 0xFFF);
+        public int MethodSubchannel => (int)((Method >> 13) & 0x7);
+        public TertOp TertOp => (TertOp)((Method >> 16) & 0x3);
+        public int MethodCountOld => (int)((Method >> 18) & 0x7FF);
+        public int MethodCount => (int)((Method >> 16) & 0x1FFF);
+        public int ImmdData => (int)((Method >> 16) & 0x1FFF);
+        public SecOp SecOp => (SecOp)((Method >> 29) & 0x7);
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPEntry.cs b/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPEntry.cs
new file mode 100644
index 00000000..b1b236e7
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPEntry.cs
@@ -0,0 +1,55 @@
+// This file was auto-generated from NVIDIA official Maxwell definitions.
+
+namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
+{
+    enum Entry0Fetch
+    {
+        Unconditional = 0,
+        Conditional = 1,
+    }
+
+    enum Entry1Priv
+    {
+        User = 0,
+        Kernel = 1,
+    }
+
+    enum Entry1Level
+    {
+        Main = 0,
+        Subroutine = 1,
+    }
+
+    enum Entry1Sync
+    {
+        Proceed = 0,
+        Wait = 1,
+    }
+
+    enum Entry1Opcode
+    {
+        Nop = 0,
+        Illegal = 1,
+        Crc = 2,
+        PbCrc = 3,
+    }
+
+    struct GPEntry
+    {
+#pragma warning disable CS0649
+        public uint Entry0;
+#pragma warning restore CS0649
+        public Entry0Fetch Entry0Fetch => (Entry0Fetch)((Entry0 >> 0) & 0x1);
+        public int Entry0Get => (int)((Entry0 >> 2) & 0x3FFFFFFF);
+        public int Entry0Operand => (int)(Entry0);
+#pragma warning disable CS0649
+        public uint Entry1;
+#pragma warning restore CS0649
+        public int Entry1GetHi => (int)((Entry1 >> 0) & 0xFF);
+        public Entry1Priv Entry1Priv => (Entry1Priv)((Entry1 >> 8) & 0x1);
+        public Entry1Level Entry1Level => (Entry1Level)((Entry1 >> 9) & 0x1);
+        public int Entry1Length => (int)((Entry1 >> 10) & 0x1FFFFF);
+        public Entry1Sync Entry1Sync => (Entry1Sync)((Entry1 >> 31) & 0x1);
+        public Entry1Opcode Entry1Opcode => (Entry1Opcode)((Entry1 >> 0) & 0xFF);
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoClass.cs b/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoClass.cs
new file mode 100644
index 00000000..e80d98a1
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoClass.cs
@@ -0,0 +1,248 @@
+using Ryujinx.Graphics.Device;
+using Ryujinx.Graphics.Gpu.Engine.MME;
+using System;
+using System.Collections.Generic;
+using System.Threading;
+
+namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
+{
+    /// <summary>
+    /// Represents a GPU General Purpose FIFO class.
+    /// </summary>
+    class GPFifoClass : IDeviceState
+    {
+        private readonly GpuContext _context;
+        private readonly GPFifoProcessor _parent;
+        private readonly DeviceState<GPFifoClassState> _state;
+
+        private int _previousSubChannel;
+        private bool _createSyncPending;
+
+        private const int MacrosCount = 0x80;
+
+        // Note: The size of the macro memory is unknown, we just make
+        // a guess here and use 256kb as the size. Increase if needed.
+        private const int MacroCodeSize = 256 * 256;
+
+        private readonly Macro[] _macros;
+        private readonly int[] _macroCode;
+
+        /// <summary>
+        /// Creates a new instance of the GPU General Purpose FIFO class.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        /// <param name="parent">Parent GPU General Purpose FIFO processor</param>
+        public GPFifoClass(GpuContext context, GPFifoProcessor parent)
+        {
+            _context = context;
+            _parent = parent;
+            _state = new DeviceState<GPFifoClassState>(new Dictionary<string, RwCallback>
+            {
+                { nameof(GPFifoClassState.Semaphored), new RwCallback(Semaphored, null) },
+                { nameof(GPFifoClassState.Syncpointb), new RwCallback(Syncpointb, null) },
+                { nameof(GPFifoClassState.WaitForIdle), new RwCallback(WaitForIdle, null) },
+                { nameof(GPFifoClassState.SetReference), new RwCallback(SetReference, null) },
+                { nameof(GPFifoClassState.LoadMmeInstructionRam), new RwCallback(LoadMmeInstructionRam, null) },
+                { nameof(GPFifoClassState.LoadMmeStartAddressRam), new RwCallback(LoadMmeStartAddressRam, null) },
+                { nameof(GPFifoClassState.SetMmeShadowRamControl), new RwCallback(SetMmeShadowRamControl, null) }
+            });
+
+            _macros = new Macro[MacrosCount];
+            _macroCode = new int[MacroCodeSize];
+        }
+
+        /// <summary>
+        /// Create any syncs from WaitForIdle command that are currently pending.
+        /// </summary>
+        public void CreatePendingSyncs()
+        {
+            if (_createSyncPending)
+            {
+                _createSyncPending = false;
+                _context.CreateHostSyncIfNeeded(false, false);
+            }
+        }
+
+        /// <summary>
+        /// Reads data from the class registers.
+        /// </summary>
+        /// <param name="offset">Register byte offset</param>
+        /// <returns>Data at the specified offset</returns>
+        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>
+        public void Write(int offset, int data) => _state.Write(offset, data);
+
+        /// <summary>
+        /// Writes a GPU counter to guest memory.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void Semaphored(int argument)
+        {
+            ulong address = ((ulong)_state.State.SemaphorebOffsetLower << 2) |
+                            ((ulong)_state.State.SemaphoreaOffsetUpper << 32);
+
+            int value = _state.State.SemaphorecPayload;
+
+            SemaphoredOperation operation = _state.State.SemaphoredOperation;
+
+            if (_state.State.SemaphoredReleaseSize == SemaphoredReleaseSize.SixteenBytes)
+            {
+                _parent.MemoryManager.Write(address + 4, 0);
+                _parent.MemoryManager.Write(address + 8, _context.GetTimestamp());
+            }
+
+            // TODO: Acquire operations (Wait), interrupts for invalid combinations.
+            if (operation == SemaphoredOperation.Release)
+            {
+                _parent.MemoryManager.Write(address, value);
+            }
+            else if (operation == SemaphoredOperation.Reduction)
+            {
+                bool signed = _state.State.SemaphoredFormat == SemaphoredFormat.Signed;
+
+                int mem = _parent.MemoryManager.Read<int>(address);
+
+                switch (_state.State.SemaphoredReduction)
+                {
+                    case SemaphoredReduction.Min:
+                        value = signed ? Math.Min(mem, value) : (int)Math.Min((uint)mem, (uint)value);
+                        break;
+                    case SemaphoredReduction.Max:
+                        value = signed ? Math.Max(mem, value) : (int)Math.Max((uint)mem, (uint)value);
+                        break;
+                    case SemaphoredReduction.Xor:
+                        value ^= mem;
+                        break;
+                    case SemaphoredReduction.And:
+                        value &= mem;
+                        break;
+                    case SemaphoredReduction.Or:
+                        value |= mem;
+                        break;
+                    case SemaphoredReduction.Add:
+                        value += mem;
+                        break;
+                    case SemaphoredReduction.Inc:
+                        value = (uint)mem < (uint)value ? mem + 1 : 0;
+                        break;
+                    case SemaphoredReduction.Dec:
+                        value = (uint)mem > 0 && (uint)mem <= (uint)value ? mem - 1 : value;
+                        break;
+                }
+
+                _parent.MemoryManager.Write(address, value);
+            }
+        }
+
+        /// <summary>
+        /// Apply a fence operation on a syncpoint.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void Syncpointb(int argument)
+        {
+            SyncpointbOperation operation = _state.State.SyncpointbOperation;
+
+            uint syncpointId = (uint)_state.State.SyncpointbSyncptIndex;
+
+            if (operation == SyncpointbOperation.Wait)
+            {
+                uint threshold = (uint)_state.State.SyncpointaPayload;
+
+                _context.Synchronization.WaitOnSyncpoint(syncpointId, threshold, Timeout.InfiniteTimeSpan);
+            }
+            else if (operation == SyncpointbOperation.Incr)
+            {
+                _context.CreateHostSyncIfNeeded(true, true);
+                _context.Synchronization.IncrementSyncpoint(syncpointId);
+            }
+
+            _context.AdvanceSequence();
+        }
+
+        /// <summary>
+        /// Waits for the GPU to be idle.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void WaitForIdle(int argument)
+        {
+            _parent.PerformDeferredDraws();
+            _context.Renderer.Pipeline.Barrier();
+
+            _createSyncPending = true;
+        }
+
+        /// <summary>
+        /// Used as an indirect data barrier on NVN. When used, access to previously written data must be coherent.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void SetReference(int argument)
+        {
+            _context.Renderer.Pipeline.CommandBufferBarrier();
+
+            _context.CreateHostSyncIfNeeded(false, true);
+        }
+
+        /// <summary>
+        /// Sends macro code/data to the MME.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void LoadMmeInstructionRam(int argument)
+        {
+            _macroCode[_state.State.LoadMmeInstructionRamPointer++] = argument;
+        }
+
+        /// <summary>
+        /// Binds a macro index to a position for the MME
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void LoadMmeStartAddressRam(int argument)
+        {
+            _macros[_state.State.LoadMmeStartAddressRamPointer++] = new Macro(argument);
+        }
+
+        /// <summary>
+        /// Changes the shadow RAM control.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void SetMmeShadowRamControl(int argument)
+        {
+            _parent.SetShadowRamControl(argument);
+        }
+
+        /// <summary>
+        /// Pushes an argument to a macro.
+        /// </summary>
+        /// <param name="index">Index of the macro</param>
+        /// <param name="gpuVa">GPU virtual address where the command word is located</param>
+        /// <param name="argument">Argument to be pushed to the macro</param>
+        public void MmePushArgument(int index, ulong gpuVa, int argument)
+        {
+            _macros[index].PushArgument(gpuVa, argument);
+        }
+
+        /// <summary>
+        /// Prepares a macro for execution.
+        /// </summary>
+        /// <param name="index">Index of the macro</param>
+        /// <param name="argument">Initial argument passed to the macro</param>
+        public void MmeStart(int index, int argument)
+        {
+            _macros[index].StartExecution(_context, _parent, _macroCode, argument);
+        }
+
+        /// <summary>
+        /// Executes a macro.
+        /// </summary>
+        /// <param name="index">Index of the macro</param>
+        /// <param name="state">Current GPU state</param>
+        public void CallMme(int index, IDeviceState state)
+        {
+            _macros[index].Execute(_macroCode, state);
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoClassState.cs b/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoClassState.cs
new file mode 100644
index 00000000..07d062eb
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoClassState.cs
@@ -0,0 +1,233 @@
+// This file was auto-generated from NVIDIA official Maxwell definitions.
+
+using Ryujinx.Common.Memory;
+
+namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
+{
+    /// <summary>
+    /// Semaphore operation.
+    /// </summary>
+    enum SemaphoredOperation
+    {
+        Acquire = 1,
+        Release = 2,
+        AcqGeq = 4,
+        AcqAnd = 8,
+        Reduction = 16
+    }
+
+    /// <summary>
+    /// Semaphore acquire switch enable.
+    /// </summary>
+    enum SemaphoredAcquireSwitch
+    {
+        Disabled = 0,
+        Enabled = 1
+    }
+
+    /// <summary>
+    /// Semaphore release interrupt wait enable.
+    /// </summary>
+    enum SemaphoredReleaseWfi
+    {
+        En = 0,
+        Dis = 1
+    }
+
+    /// <summary>
+    /// Semaphore release structure size.
+    /// </summary>
+    enum SemaphoredReleaseSize
+    {
+        SixteenBytes = 0,
+        FourBytes = 1
+    }
+
+    /// <summary>
+    /// Semaphore reduction operation.
+    /// </summary>
+    enum SemaphoredReduction
+    {
+        Min = 0,
+        Max = 1,
+        Xor = 2,
+        And = 3,
+        Or = 4,
+        Add = 5,
+        Inc = 6,
+        Dec = 7
+    }
+
+    /// <summary>
+    /// Semaphore format.
+    /// </summary>
+    enum SemaphoredFormat
+    {
+        Signed = 0,
+        Unsigned = 1
+    }
+
+    /// <summary>
+    /// Memory Translation Lookaside Buffer Page Directory Buffer invalidation.
+    /// </summary>
+    enum MemOpCTlbInvalidatePdb
+    {
+        One = 0,
+        All = 1
+    }
+
+    /// <summary>
+    /// Memory Translation Lookaside Buffer GPC invalidation enable.
+    /// </summary>
+    enum MemOpCTlbInvalidateGpc
+    {
+        Enable = 0,
+        Disable = 1
+    }
+
+    /// <summary>
+    /// Memory Translation Lookaside Buffer invalidation target.
+    /// </summary>
+    enum MemOpCTlbInvalidateTarget
+    {
+        VidMem = 0,
+        SysMemCoherent = 2,
+        SysMemNoncoherent = 3
+    }
+
+    /// <summary>
+    /// Memory operation.
+    /// </summary>
+    enum MemOpDOperation
+    {
+        Membar = 5,
+        MmuTlbInvalidate = 9,
+        L2PeermemInvalidate = 13,
+        L2SysmemInvalidate = 14,
+        L2CleanComptags = 15,
+        L2FlushDirty = 16
+    }
+
+    /// <summary>
+    /// Syncpoint operation.
+    /// </summary>
+    enum SyncpointbOperation
+    {
+        Wait = 0,
+        Incr = 1
+    }
+
+    /// <summary>
+    /// Syncpoint wait switch enable.
+    /// </summary>
+    enum SyncpointbWaitSwitch
+    {
+        Dis = 0,
+        En = 1
+    }
+
+    /// <summary>
+    /// Wait for interrupt scope.
+    /// </summary>
+    enum WfiScope
+    {
+        CurrentScgType = 0,
+        All = 1
+    }
+
+    /// <summary>
+    /// Yield operation.
+    /// </summary>
+    enum YieldOp
+    {
+        Nop = 0,
+        PbdmaTimeslice = 1,
+        RunlistTimeslice = 2,
+        Tsg = 3
+    }
+
+    /// <summary>
+    /// General Purpose FIFO class state.
+    /// </summary>
+    struct GPFifoClassState
+    {
+#pragma warning disable CS0649
+        public uint SetObject;
+        public int SetObjectNvclass => (int)((SetObject >> 0) & 0xFFFF);
+        public int SetObjectEngine => (int)((SetObject >> 16) & 0x1F);
+        public uint Illegal;
+        public int IllegalHandle => (int)(Illegal);
+        public uint Nop;
+        public int NopHandle => (int)(Nop);
+        public uint Reserved0C;
+        public uint Semaphorea;
+        public int SemaphoreaOffsetUpper => (int)((Semaphorea >> 0) & 0xFF);
+        public uint Semaphoreb;
+        public int SemaphorebOffsetLower => (int)((Semaphoreb >> 2) & 0x3FFFFFFF);
+        public uint Semaphorec;
+        public int SemaphorecPayload => (int)(Semaphorec);
+        public uint Semaphored;
+        public SemaphoredOperation SemaphoredOperation => (SemaphoredOperation)((Semaphored >> 0) & 0x1F);
+        public SemaphoredAcquireSwitch SemaphoredAcquireSwitch => (SemaphoredAcquireSwitch)((Semaphored >> 12) & 0x1);
+        public SemaphoredReleaseWfi SemaphoredReleaseWfi => (SemaphoredReleaseWfi)((Semaphored >> 20) & 0x1);
+        public SemaphoredReleaseSize SemaphoredReleaseSize => (SemaphoredReleaseSize)((Semaphored >> 24) & 0x1);
+        public SemaphoredReduction SemaphoredReduction => (SemaphoredReduction)((Semaphored >> 27) & 0xF);
+        public SemaphoredFormat SemaphoredFormat => (SemaphoredFormat)((Semaphored >> 31) & 0x1);
+        public uint NonStallInterrupt;
+        public int NonStallInterruptHandle => (int)(NonStallInterrupt);
+        public uint FbFlush;
+        public int FbFlushHandle => (int)(FbFlush);
+        public uint Reserved28;
+        public uint Reserved2C;
+        public uint MemOpC;
+        public int MemOpCOperandLow => (int)((MemOpC >> 2) & 0x3FFFFFFF);
+        public MemOpCTlbInvalidatePdb MemOpCTlbInvalidatePdb => (MemOpCTlbInvalidatePdb)((MemOpC >> 0) & 0x1);
+        public MemOpCTlbInvalidateGpc MemOpCTlbInvalidateGpc => (MemOpCTlbInvalidateGpc)((MemOpC >> 1) & 0x1);
+        public MemOpCTlbInvalidateTarget MemOpCTlbInvalidateTarget => (MemOpCTlbInvalidateTarget)((MemOpC >> 10) & 0x3);
+        public int MemOpCTlbInvalidateAddrLo => (int)((MemOpC >> 12) & 0xFFFFF);
+        public uint MemOpD;
+        public int MemOpDOperandHigh => (int)((MemOpD >> 0) & 0xFF);
+        public MemOpDOperation MemOpDOperation => (MemOpDOperation)((MemOpD >> 27) & 0x1F);
+        public int MemOpDTlbInvalidateAddrHi => (int)((MemOpD >> 0) & 0xFF);
+        public uint Reserved38;
+        public uint Reserved3C;
+        public uint Reserved40;
+        public uint Reserved44;
+        public uint Reserved48;
+        public uint Reserved4C;
+        public uint SetReference;
+        public int SetReferenceCount => (int)(SetReference);
+        public uint Reserved54;
+        public uint Reserved58;
+        public uint Reserved5C;
+        public uint Reserved60;
+        public uint Reserved64;
+        public uint Reserved68;
+        public uint Reserved6C;
+        public uint Syncpointa;
+        public int SyncpointaPayload => (int)(Syncpointa);
+        public uint Syncpointb;
+        public SyncpointbOperation SyncpointbOperation => (SyncpointbOperation)((Syncpointb >> 0) & 0x1);
+        public SyncpointbWaitSwitch SyncpointbWaitSwitch => (SyncpointbWaitSwitch)((Syncpointb >> 4) & 0x1);
+        public int SyncpointbSyncptIndex => (int)((Syncpointb >> 8) & 0xFFF);
+        public uint Wfi;
+        public WfiScope WfiScope => (WfiScope)((Wfi >> 0) & 0x1);
+        public uint CrcCheck;
+        public int CrcCheckValue => (int)(CrcCheck);
+        public uint Yield;
+        public YieldOp YieldOp => (YieldOp)((Yield >> 0) & 0x3);
+        // TODO: Eventually move this to per-engine state.
+        public Array31<uint> Reserved84;
+        public uint NoOperation;
+        public uint SetNotifyA;
+        public uint SetNotifyB;
+        public uint Notify;
+        public uint WaitForIdle;
+        public uint LoadMmeInstructionRamPointer;
+        public uint LoadMmeInstructionRam;
+        public uint LoadMmeStartAddressRamPointer;
+        public uint LoadMmeStartAddressRam;
+        public uint SetMmeShadowRamControl;
+#pragma warning restore CS0649
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoDevice.cs b/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoDevice.cs
new file mode 100644
index 00000000..cd29a9da
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoDevice.cs
@@ -0,0 +1,262 @@
+using Ryujinx.Graphics.Gpu.Memory;
+using System;
+using System.Collections.Concurrent;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+using System.Threading;
+
+namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
+{
+    /// <summary>
+    /// Represents a GPU General Purpose FIFO device.
+    /// </summary>
+    public sealed class GPFifoDevice : IDisposable
+    {
+        /// <summary>
+        /// Indicates if the command buffer has pre-fetch enabled.
+        /// </summary>
+        private enum CommandBufferType
+        {
+            Prefetch,
+            NoPrefetch
+        }
+
+        /// <summary>
+        /// Command buffer data.
+        /// </summary>
+        private struct CommandBuffer
+        {
+            /// <summary>
+            /// Processor used to process the command buffer. Contains channel state.
+            /// </summary>
+            public GPFifoProcessor Processor;
+
+            /// <summary>
+            /// The type of the command buffer.
+            /// </summary>
+            public CommandBufferType Type;
+
+            /// <summary>
+            /// Fetched data.
+            /// </summary>
+            public int[] Words;
+
+            /// <summary>
+            /// The GPFIFO entry address (used in <see cref="CommandBufferType.NoPrefetch"/> mode).
+            /// </summary>
+            public ulong EntryAddress;
+
+            /// <summary>
+            /// The count of entries inside this GPFIFO entry.
+            /// </summary>
+            public uint EntryCount;
+
+            /// <summary>
+            /// Get the entries for the command buffer from memory.
+            /// </summary>
+            /// <param name="memoryManager">The memory manager used to fetch the data</param>
+            /// <param name="flush">If true, flushes potential GPU written data before reading the command buffer</param>
+            /// <returns>The fetched data</returns>
+            private ReadOnlySpan<int> GetWords(MemoryManager memoryManager, bool flush)
+            {
+                return MemoryMarshal.Cast<byte, int>(memoryManager.GetSpan(EntryAddress, (int)EntryCount * 4, flush));
+            }
+
+            /// <summary>
+            /// Prefetch the command buffer.
+            /// </summary>
+            /// <param name="memoryManager">The memory manager used to fetch the data</param>
+            public void Prefetch(MemoryManager memoryManager)
+            {
+                Words = GetWords(memoryManager, true).ToArray();
+            }
+
+            /// <summary>
+            /// Fetch the command buffer.
+            /// </summary>
+            /// <param name="memoryManager">The memory manager used to fetch the data</param>
+            /// <param name="flush">If true, flushes potential GPU written data before reading the command buffer</param>
+            /// <returns>The command buffer words</returns>
+            public ReadOnlySpan<int> Fetch(MemoryManager memoryManager, bool flush)
+            {
+                return Words ?? GetWords(memoryManager, flush);
+            }
+        }
+
+        private readonly ConcurrentQueue<CommandBuffer> _commandBufferQueue;
+
+        private CommandBuffer _currentCommandBuffer;
+        private GPFifoProcessor _prevChannelProcessor;
+
+        private readonly bool _ibEnable;
+        private readonly GpuContext _context;
+        private readonly AutoResetEvent _event;
+
+        private bool _interrupt;
+        private int _flushSkips;
+
+        /// <summary>
+        /// Creates a new instance of the GPU General Purpose FIFO device.
+        /// </summary>
+        /// <param name="context">GPU context that the GPFIFO belongs to</param>
+        internal GPFifoDevice(GpuContext context)
+        {
+            _commandBufferQueue = new ConcurrentQueue<CommandBuffer>();
+            _ibEnable = true;
+            _context = context;
+            _event = new AutoResetEvent(false);
+        }
+
+        /// <summary>
+        /// Signal the FIFO that there are new entries to process.
+        /// </summary>
+        public void SignalNewEntries()
+        {
+            _event.Set();
+        }
+
+        /// <summary>
+        /// Push a GPFIFO entry in the form of a prefetched command buffer.
+        /// It is intended to be used by nvservices to handle special cases.
+        /// </summary>
+        /// <param name="processor">Processor used to process <paramref name="commandBuffer"/></param>
+        /// <param name="commandBuffer">The command buffer containing the prefetched commands</param>
+        internal void PushHostCommandBuffer(GPFifoProcessor processor, int[] commandBuffer)
+        {
+            _commandBufferQueue.Enqueue(new CommandBuffer
+            {
+                Processor = processor,
+                Type = CommandBufferType.Prefetch,
+                Words = commandBuffer,
+                EntryAddress = ulong.MaxValue,
+                EntryCount = (uint)commandBuffer.Length
+            });
+        }
+
+        /// <summary>
+        /// Create a CommandBuffer from a GPFIFO entry.
+        /// </summary>
+        /// <param name="processor">Processor used to process the command buffer pointed to by <paramref name="entry"/></param>
+        /// <param name="entry">The GPFIFO entry</param>
+        /// <returns>A new CommandBuffer based on the GPFIFO entry</returns>
+        private static CommandBuffer CreateCommandBuffer(GPFifoProcessor processor, GPEntry entry)
+        {
+            CommandBufferType type = CommandBufferType.Prefetch;
+
+            if (entry.Entry1Sync == Entry1Sync.Wait)
+            {
+                type = CommandBufferType.NoPrefetch;
+            }
+
+            ulong startAddress = ((ulong)entry.Entry0Get << 2) | ((ulong)entry.Entry1GetHi << 32);
+
+            return new CommandBuffer
+            {
+                Processor = processor,
+                Type = type,
+                Words = null,
+                EntryAddress = startAddress,
+                EntryCount = (uint)entry.Entry1Length
+            };
+        }
+
+        /// <summary>
+        /// Pushes GPFIFO entries.
+        /// </summary>
+        /// <param name="processor">Processor used to process the command buffers pointed to by <paramref name="entries"/></param>
+        /// <param name="entries">GPFIFO entries</param>
+        internal void PushEntries(GPFifoProcessor processor, ReadOnlySpan<ulong> entries)
+        {
+            bool beforeBarrier = true;
+
+            for (int index = 0; index < entries.Length; index++)
+            {
+                ulong entry = entries[index];
+
+                CommandBuffer commandBuffer = CreateCommandBuffer(processor, Unsafe.As<ulong, GPEntry>(ref entry));
+
+                if (beforeBarrier && commandBuffer.Type == CommandBufferType.Prefetch)
+                {
+                    commandBuffer.Prefetch(processor.MemoryManager);
+                }
+
+                if (commandBuffer.Type == CommandBufferType.NoPrefetch)
+                {
+                    beforeBarrier = false;
+                }
+
+                _commandBufferQueue.Enqueue(commandBuffer);
+            }
+        }
+
+        /// <summary>
+        /// Waits until commands are pushed to the FIFO.
+        /// </summary>
+        /// <returns>True if commands were received, false if wait timed out</returns>
+        public bool WaitForCommands()
+        {
+            return !_commandBufferQueue.IsEmpty || (_event.WaitOne(8) && !_commandBufferQueue.IsEmpty);
+        }
+
+        /// <summary>
+        /// Processes commands pushed to the FIFO.
+        /// </summary>
+        public void DispatchCalls()
+        {
+            // Use this opportunity to also dispose any pending channels that were closed.
+            _context.RunDeferredActions();
+
+            // Process command buffers.
+            while (_ibEnable && !_interrupt && _commandBufferQueue.TryDequeue(out CommandBuffer entry))
+            {
+                bool flushCommandBuffer = true;
+
+                if (_flushSkips != 0)
+                {
+                    _flushSkips--;
+                    flushCommandBuffer = false;
+                }
+
+                _currentCommandBuffer = entry;
+                ReadOnlySpan<int> words = entry.Fetch(entry.Processor.MemoryManager, flushCommandBuffer);
+
+                // If we are changing the current channel,
+                // we need to force all the host state to be updated.
+                if (_prevChannelProcessor != entry.Processor)
+                {
+                    _prevChannelProcessor = entry.Processor;
+                    entry.Processor.ForceAllDirty();
+                }
+
+                entry.Processor.Process(entry.EntryAddress, words);
+            }
+
+            _interrupt = false;
+        }
+
+        /// <summary>
+        /// Sets the number of flushes that should be skipped for subsequent command buffers.
+        /// </summary>
+        /// <remarks>
+        /// This can improve performance when command buffer data only needs to be consumed by the GPU.
+        /// </remarks>
+        /// <param name="count">The amount of flushes that should be skipped</param>
+        internal void SetFlushSkips(int count)
+        {
+            _flushSkips = count;
+        }
+
+        /// <summary>
+        /// Interrupts command processing. This will break out of the DispatchCalls loop.
+        /// </summary>
+        public void Interrupt()
+        {
+            _interrupt = true;
+        }
+
+        /// <summary>
+        /// Disposes of resources used for GPFifo command processing.
+        /// </summary>
+        public void Dispose() => _event.Dispose();
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoProcessor.cs b/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoProcessor.cs
new file mode 100644
index 00000000..3fb3feee
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/GPFifo/GPFifoProcessor.cs
@@ -0,0 +1,331 @@
+using Ryujinx.Graphics.Device;
+using Ryujinx.Graphics.Gpu.Engine.Compute;
+using Ryujinx.Graphics.Gpu.Engine.Dma;
+using Ryujinx.Graphics.Gpu.Engine.InlineToMemory;
+using Ryujinx.Graphics.Gpu.Engine.Threed;
+using Ryujinx.Graphics.Gpu.Engine.Twod;
+using Ryujinx.Graphics.Gpu.Memory;
+using System;
+using System.Runtime.CompilerServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.GPFifo
+{
+    /// <summary>
+    /// Represents a GPU General Purpose FIFO command processor.
+    /// </summary>
+    class GPFifoProcessor
+    {
+        private const int MacrosCount = 0x80;
+        private const int MacroIndexMask = MacrosCount - 1;
+
+        private const int LoadInlineDataMethodOffset = 0x6d;
+        private const int UniformBufferUpdateDataMethodOffset = 0x8e4;
+
+        private readonly GpuChannel _channel;
+
+        /// <summary>
+        /// Channel memory manager.
+        /// </summary>
+        public MemoryManager MemoryManager => _channel.MemoryManager;
+
+        /// <summary>
+        /// 3D Engine.
+        /// </summary>
+        public ThreedClass ThreedClass => _3dClass;
+
+        /// <summary>
+        /// Internal GPFIFO state.
+        /// </summary>
+        private struct DmaState
+        {
+            public int Method;
+            public int SubChannel;
+            public int MethodCount;
+            public bool NonIncrementing;
+            public bool IncrementOnce;
+        }
+
+        private DmaState _state;
+
+        private readonly ThreedClass _3dClass;
+        private readonly ComputeClass _computeClass;
+        private readonly InlineToMemoryClass _i2mClass;
+        private readonly TwodClass _2dClass;
+        private readonly DmaClass _dmaClass;
+
+        private readonly GPFifoClass _fifoClass;
+
+        /// <summary>
+        /// Creates a new instance of the GPU General Purpose FIFO command processor.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        /// <param name="channel">Channel that the GPFIFO processor belongs to</param>
+        public GPFifoProcessor(GpuContext context, GpuChannel channel)
+        {
+            _channel = channel;
+
+            _fifoClass = new GPFifoClass(context, this);
+            _3dClass = new ThreedClass(context, channel, _fifoClass);
+            _computeClass = new ComputeClass(context, channel, _3dClass);
+            _i2mClass = new InlineToMemoryClass(context, channel);
+            _2dClass = new TwodClass(channel);
+            _dmaClass = new DmaClass(context, channel, _3dClass);
+        }
+
+        /// <summary>
+        /// Processes a command buffer.
+        /// </summary>
+        /// <param name="baseGpuVa">Base GPU virtual address of the command buffer</param>
+        /// <param name="commandBuffer">Command buffer</param>
+        public void Process(ulong baseGpuVa, ReadOnlySpan<int> commandBuffer)
+        {
+            for (int index = 0; index < commandBuffer.Length; index++)
+            {
+                int command = commandBuffer[index];
+
+                ulong gpuVa = baseGpuVa + (ulong)index * 4;
+
+                if (_state.MethodCount != 0)
+                {
+                    if (TryFastI2mBufferUpdate(commandBuffer, ref index))
+                    {
+                        continue;
+                    }
+
+                    Send(gpuVa, _state.Method, command, _state.SubChannel, _state.MethodCount <= 1);
+
+                    if (!_state.NonIncrementing)
+                    {
+                        _state.Method++;
+                    }
+
+                    if (_state.IncrementOnce)
+                    {
+                        _state.NonIncrementing = true;
+                    }
+
+                    _state.MethodCount--;
+                }
+                else
+                {
+                    CompressedMethod meth = Unsafe.As<int, CompressedMethod>(ref command);
+
+                    if (TryFastUniformBufferUpdate(meth, commandBuffer, index))
+                    {
+                        index += meth.MethodCount;
+                        continue;
+                    }
+
+                    switch (meth.SecOp)
+                    {
+                        case SecOp.IncMethod:
+                        case SecOp.NonIncMethod:
+                        case SecOp.OneInc:
+                            _state.Method = meth.MethodAddress;
+                            _state.SubChannel = meth.MethodSubchannel;
+                            _state.MethodCount = meth.MethodCount;
+                            _state.IncrementOnce = meth.SecOp == SecOp.OneInc;
+                            _state.NonIncrementing = meth.SecOp == SecOp.NonIncMethod;
+                            break;
+                        case SecOp.ImmdDataMethod:
+                            Send(gpuVa, meth.MethodAddress, meth.ImmdData, meth.MethodSubchannel, true);
+                            break;
+                    }
+                }
+            }
+
+            _3dClass.FlushUboDirty();
+        }
+
+        /// <summary>
+        /// Tries to perform a fast Inline-to-Memory data update.
+        /// If successful, all data will be copied at once, and <see cref="DmaState.MethodCount"/>
+        /// command buffer entries will be consumed.
+        /// </summary>
+        /// <param name="commandBuffer">Command buffer where the data is contained</param>
+        /// <param name="offset">Offset at <paramref name="commandBuffer"/> where the data is located, auto-incremented on success</param>
+        /// <returns>True if the fast copy was successful, false otherwise</returns>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private bool TryFastI2mBufferUpdate(ReadOnlySpan<int> commandBuffer, ref int offset)
+        {
+            if (_state.Method == LoadInlineDataMethodOffset && _state.NonIncrementing && _state.SubChannel <= 2)
+            {
+                int availableCount = commandBuffer.Length - offset;
+                int consumeCount = Math.Min(_state.MethodCount, availableCount);
+
+                var data = commandBuffer.Slice(offset, consumeCount);
+
+                if (_state.SubChannel == 0)
+                {
+                    _3dClass.LoadInlineData(data);
+                }
+                else if (_state.SubChannel == 1)
+                {
+                    _computeClass.LoadInlineData(data);
+                }
+                else /* if (_state.SubChannel == 2) */
+                {
+                    _i2mClass.LoadInlineData(data);
+                }
+
+                offset += consumeCount - 1;
+                _state.MethodCount -= consumeCount;
+
+                return true;
+            }
+
+            return false;
+        }
+
+        /// <summary>
+        /// Tries to perform a fast constant buffer data update.
+        /// If successful, all data will be copied at once, and <see cref="CompressedMethod.MethodCount"/> + 1
+        /// command buffer entries will be consumed.
+        /// </summary>
+        /// <param name="meth">Compressed method to be checked</param>
+        /// <param name="commandBuffer">Command buffer where <paramref name="meth"/> is contained</param>
+        /// <param name="offset">Offset at <paramref name="commandBuffer"/> where <paramref name="meth"/> is located</param>
+        /// <returns>True if the fast copy was successful, false otherwise</returns>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private bool TryFastUniformBufferUpdate(CompressedMethod meth, ReadOnlySpan<int> commandBuffer, int offset)
+        {
+            int availableCount = commandBuffer.Length - offset;
+
+            if (meth.MethodAddress == UniformBufferUpdateDataMethodOffset &&
+                meth.MethodCount < availableCount &&
+                meth.SecOp == SecOp.NonIncMethod)
+            {
+                _3dClass.ConstantBufferUpdate(commandBuffer.Slice(offset + 1, meth.MethodCount));
+
+                return true;
+            }
+
+            return false;
+        }
+
+        /// <summary>
+        /// Sends a uncompressed method for processing by the graphics pipeline.
+        /// </summary>
+        /// <param name="gpuVa">GPU virtual address where the command word is located</param>
+        /// <param name="meth">Method to be processed</param>
+        private void Send(ulong gpuVa, int offset, int argument, int subChannel, bool isLastCall)
+        {
+            if (offset < 0x60)
+            {
+                _fifoClass.Write(offset * 4, argument);
+            }
+            else if (offset < 0xe00)
+            {
+                offset *= 4;
+
+                switch (subChannel)
+                {
+                    case 0:
+                        _3dClass.Write(offset, argument);
+                        break;
+                    case 1:
+                        _computeClass.Write(offset, argument);
+                        break;
+                    case 2:
+                        _i2mClass.Write(offset, argument);
+                        break;
+                    case 3:
+                        _2dClass.Write(offset, argument);
+                        break;
+                    case 4:
+                        _dmaClass.Write(offset, argument);
+                        break;
+                }
+            }
+            else
+            {
+                IDeviceState state = subChannel switch
+                {
+                    0 => _3dClass,
+                    3 => _2dClass,
+                    _ => null
+                };
+
+                if (state != null)
+                {
+                    int macroIndex = (offset >> 1) & MacroIndexMask;
+
+                    if ((offset & 1) != 0)
+                    {
+                        _fifoClass.MmePushArgument(macroIndex, gpuVa, argument);
+                    }
+                    else
+                    {
+                        _fifoClass.MmeStart(macroIndex, argument);
+                    }
+
+                    if (isLastCall)
+                    {
+                        _fifoClass.CallMme(macroIndex, state);
+
+                        _3dClass.PerformDeferredDraws();
+                    }
+                }
+            }
+        }
+
+        /// <summary>
+        /// Writes data directly to the state of the specified class.
+        /// </summary>
+        /// <param name="classId">ID of the class to write the data into</param>
+        /// <param name="offset">State offset in bytes</param>
+        /// <param name="value">Value to be written</param>
+        public void Write(ClassId classId, int offset, int value)
+        {
+            switch (classId)
+            {
+                case ClassId.Threed:
+                    _3dClass.Write(offset, value);
+                    break;
+                case ClassId.Compute:
+                    _computeClass.Write(offset, value);
+                    break;
+                case ClassId.InlineToMemory:
+                    _i2mClass.Write(offset, value);
+                    break;
+                case ClassId.Twod:
+                    _2dClass.Write(offset, value);
+                    break;
+                case ClassId.Dma:
+                    _dmaClass.Write(offset, value);
+                    break;
+                case ClassId.GPFifo:
+                    _fifoClass.Write(offset, value);
+                    break;
+            }
+        }
+
+        /// <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)
+        {
+            _3dClass.SetShadowRamControl(control);
+        }
+
+        /// <summary>
+        /// Forces a full host state update by marking all state as modified,
+        /// and also requests all GPU resources in use to be rebound.
+        /// </summary>
+        public void ForceAllDirty()
+        {
+            _3dClass.ForceStateDirty();
+            _channel.BufferManager.Rebind();
+            _channel.TextureManager.Rebind();
+        }
+
+        /// <summary>
+        /// Perform any deferred draws.
+        /// </summary>
+        public void PerformDeferredDraws()
+        {
+            _3dClass.PerformDeferredDraws();
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/InlineToMemory/InlineToMemoryClass.cs b/src/Ryujinx.Graphics.Gpu/Engine/InlineToMemory/InlineToMemoryClass.cs
new file mode 100644
index 00000000..e1d7e940
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/InlineToMemory/InlineToMemoryClass.cs
@@ -0,0 +1,273 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.Device;
+using Ryujinx.Graphics.Texture;
+using System;
+using System.Collections.Generic;
+using System.Runtime.InteropServices;
+using System.Runtime.Intrinsics;
+
+namespace Ryujinx.Graphics.Gpu.Engine.InlineToMemory
+{
+    /// <summary>
+    /// Represents a Inline-to-Memory engine class.
+    /// </summary>
+    class InlineToMemoryClass : IDeviceState
+    {
+        private readonly GpuContext _context;
+        private readonly GpuChannel _channel;
+        private readonly DeviceState<InlineToMemoryClassState> _state;
+
+        private bool _isLinear;
+
+        private int _offset;
+        private int _size;
+
+        private ulong _dstGpuVa;
+        private int _dstX;
+        private int _dstY;
+        private int _dstWidth;
+        private int _dstHeight;
+        private int _dstStride;
+        private int _dstGobBlocksInY;
+        private int _dstGobBlocksInZ;
+        private int _lineLengthIn;
+        private int _lineCount;
+
+        private bool _finished;
+
+        private int[] _buffer;
+
+        /// <summary>
+        /// Creates a new instance of the Inline-to-Memory engine class.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        /// <param name="channel">GPU channel</param>
+        /// <param name="initializeState">Indicates if the internal state should be initialized. Set to false if part of another engine</param>
+        public InlineToMemoryClass(GpuContext context, GpuChannel channel, bool initializeState)
+        {
+            _context = context;
+            _channel = channel;
+
+            if (initializeState)
+            {
+                _state = new DeviceState<InlineToMemoryClassState>(new Dictionary<string, RwCallback>
+                {
+                    { nameof(InlineToMemoryClassState.LaunchDma), new RwCallback(LaunchDma, null) },
+                    { nameof(InlineToMemoryClassState.LoadInlineData), new RwCallback(LoadInlineData, null) }
+                });
+            }
+        }
+
+        /// <summary>
+        /// Creates a new instance of the inline-to-memory engine class.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        /// <param name="channel">GPU channel</param>
+        public InlineToMemoryClass(GpuContext context, GpuChannel channel) : this(context, channel, true)
+        {
+        }
+
+        /// <summary>
+        /// Reads data from the class registers.
+        /// </summary>
+        /// <param name="offset">Register byte offset</param>
+        /// <returns>Data at the specified offset</returns>
+        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>
+        public void Write(int offset, int data) => _state.Write(offset, data);
+
+        /// <summary>
+        /// Launches Inline-to-Memory engine DMA copy.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void LaunchDma(int argument)
+        {
+            LaunchDma(ref _state.State, argument);
+        }
+
+        /// <summary>
+        /// Launches Inline-to-Memory engine DMA copy.
+        /// </summary>
+        /// <param name="state">Current class state</param>
+        /// <param name="argument">Method call argument</param>
+        public void LaunchDma(ref InlineToMemoryClassState state, int argument)
+        {
+            _isLinear = (argument & 1) != 0;
+
+            _offset = 0;
+            _size = (int)(BitUtils.AlignUp<uint>(state.LineLengthIn, 4) * state.LineCount);
+
+            int count = _size / 4;
+
+            if (_buffer == null || _buffer.Length < count)
+            {
+                _buffer = new int[count];
+            }
+
+            ulong dstGpuVa = ((ulong)state.OffsetOutUpperValue << 32) | state.OffsetOut;
+
+            _dstGpuVa = dstGpuVa;
+            _dstX = state.SetDstOriginBytesXV;
+            _dstY = state.SetDstOriginSamplesYV;
+            _dstWidth = (int)state.SetDstWidth;
+            _dstHeight = (int)state.SetDstHeight;
+            _dstStride = (int)state.PitchOut;
+            _dstGobBlocksInY = 1 << (int)state.SetDstBlockSizeHeight;
+            _dstGobBlocksInZ = 1 << (int)state.SetDstBlockSizeDepth;
+            _lineLengthIn = (int)state.LineLengthIn;
+            _lineCount = (int)state.LineCount;
+
+            _finished = false;
+        }
+
+        /// <summary>
+        /// Pushes a block of data to the Inline-to-Memory engine.
+        /// </summary>
+        /// <param name="data">Data to push</param>
+        public void LoadInlineData(ReadOnlySpan<int> data)
+        {
+            if (!_finished)
+            {
+                int copySize = Math.Min(data.Length, _buffer.Length - _offset);
+                data.Slice(0, copySize).CopyTo(new Span<int>(_buffer).Slice(_offset, copySize));
+
+                _offset += copySize;
+
+                if (_offset * 4 >= _size)
+                {
+                    FinishTransfer();
+                }
+            }
+        }
+
+        /// <summary>
+        /// Pushes a word of data to the Inline-to-Memory engine.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void LoadInlineData(int argument)
+        {
+            if (!_finished)
+            {
+                _buffer[_offset++] = argument;
+
+                if (_offset * 4 >= _size)
+                {
+                    FinishTransfer();
+                }
+            }
+        }
+
+        /// <summary>
+        /// Performs actual copy of the inline data after the transfer is finished.
+        /// </summary>
+        private void FinishTransfer()
+        {
+            var memoryManager = _channel.MemoryManager;
+
+            var data = MemoryMarshal.Cast<int, byte>(_buffer).Slice(0, _size);
+
+            if (_isLinear && _lineCount == 1)
+            {
+                memoryManager.WriteTrackedResource(_dstGpuVa, data.Slice(0, _lineLengthIn));
+                _context.AdvanceSequence();
+            }
+            else
+            {
+                // TODO: Verify if the destination X/Y and width/height are taken into account
+                // for linear texture transfers. If not, we can use the fast path for that aswell.
+                // Right now the copy code at the bottom assumes that it is used on both which might be incorrect.
+                if (!_isLinear)
+                {
+                    var target = memoryManager.Physical.TextureCache.FindTexture(
+                        memoryManager,
+                        _dstGpuVa,
+                        1,
+                        _dstStride,
+                        _dstHeight,
+                        _lineLengthIn,
+                        _lineCount,
+                        _isLinear,
+                        _dstGobBlocksInY,
+                        _dstGobBlocksInZ);
+
+                    if (target != null)
+                    {
+                        target.SynchronizeMemory();
+                        target.SetData(data, 0, 0, new GAL.Rectangle<int>(_dstX, _dstY, _lineLengthIn / target.Info.FormatInfo.BytesPerPixel, _lineCount));
+                        target.SignalModified();
+
+                        return;
+                    }
+                }
+
+                var dstCalculator = new OffsetCalculator(
+                    _dstWidth,
+                    _dstHeight,
+                    _dstStride,
+                    _isLinear,
+                    _dstGobBlocksInY,
+                    1);
+
+                int srcOffset = 0;
+
+                for (int y = _dstY; y < _dstY + _lineCount; y++)
+                {
+                    int x1 = _dstX;
+                    int x2 = _dstX + _lineLengthIn;
+                    int x1Round = BitUtils.AlignUp(_dstX, 16);
+                    int x2Trunc = BitUtils.AlignDown(x2, 16);
+
+                    int x = x1;
+
+                    if (x1Round <= x2)
+                    {
+                        for (; x < x1Round; x++, srcOffset++)
+                        {
+                            int dstOffset = dstCalculator.GetOffset(x, y);
+
+                            ulong dstAddress = _dstGpuVa + (uint)dstOffset;
+
+                            memoryManager.Write(dstAddress, data[srcOffset]);
+                        }
+                    }
+
+                    for (; x < x2Trunc; x += 16, srcOffset += 16)
+                    {
+                        int dstOffset = dstCalculator.GetOffset(x, y);
+
+                        ulong dstAddress = _dstGpuVa + (uint)dstOffset;
+
+                        memoryManager.Write(dstAddress, MemoryMarshal.Cast<byte, Vector128<byte>>(data.Slice(srcOffset, 16))[0]);
+                    }
+
+                    for (; x < x2; x++, srcOffset++)
+                    {
+                        int dstOffset = dstCalculator.GetOffset(x, y);
+
+                        ulong dstAddress = _dstGpuVa + (uint)dstOffset;
+
+                        memoryManager.Write(dstAddress, data[srcOffset]);
+                    }
+
+                    // All lines must be aligned to 4 bytes, as the data is pushed one word at a time.
+                    // If our copy length is not a multiple of 4, then we need to skip the padding bytes here.
+                    int misalignment = _lineLengthIn & 3;
+
+                    if (misalignment != 0)
+                    {
+                        srcOffset += 4 - misalignment;
+                    }
+                }
+
+                _context.AdvanceSequence();
+            }
+
+            _finished = true;
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/InlineToMemory/InlineToMemoryClassState.cs b/src/Ryujinx.Graphics.Gpu/Engine/InlineToMemory/InlineToMemoryClassState.cs
new file mode 100644
index 00000000..d0c82a5e
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/InlineToMemory/InlineToMemoryClassState.cs
@@ -0,0 +1,181 @@
+// This file was auto-generated from NVIDIA official Maxwell definitions.
+
+namespace Ryujinx.Graphics.Gpu.Engine.InlineToMemory
+{
+    /// <summary>
+    /// Notify type.
+    /// </summary>
+    enum NotifyType
+    {
+        WriteOnly = 0,
+        WriteThenAwaken = 1,
+    }
+
+    /// <summary>
+    /// Width in GOBs of the destination texture.
+    /// </summary>
+    enum SetDstBlockSizeWidth
+    {
+        OneGob = 0,
+    }
+
+    /// <summary>
+    /// Height in GOBs of the destination texture.
+    /// </summary>
+    enum SetDstBlockSizeHeight
+    {
+        OneGob = 0,
+        TwoGobs = 1,
+        FourGobs = 2,
+        EightGobs = 3,
+        SixteenGobs = 4,
+        ThirtytwoGobs = 5,
+    }
+
+    /// <summary>
+    /// Depth in GOBs of the destination texture.
+    /// </summary>
+    enum SetDstBlockSizeDepth
+    {
+        OneGob = 0,
+        TwoGobs = 1,
+        FourGobs = 2,
+        EightGobs = 3,
+        SixteenGobs = 4,
+        ThirtytwoGobs = 5,
+    }
+
+    /// <summary>
+    /// Memory layout of the destination texture.
+    /// </summary>
+    enum LaunchDmaDstMemoryLayout
+    {
+        Blocklinear = 0,
+        Pitch = 1,
+    }
+
+    /// <summary>
+    /// DMA completion type.
+    /// </summary>
+    enum LaunchDmaCompletionType
+    {
+        FlushDisable = 0,
+        FlushOnly = 1,
+        ReleaseSemaphore = 2,
+    }
+
+    /// <summary>
+    /// DMA interrupt type.
+    /// </summary>
+    enum LaunchDmaInterruptType
+    {
+        None = 0,
+        Interrupt = 1,
+    }
+
+    /// <summary>
+    /// DMA semaphore structure size.
+    /// </summary>
+    enum LaunchDmaSemaphoreStructSize
+    {
+        FourWords = 0,
+        OneWord = 1,
+    }
+
+    /// <summary>
+    /// DMA semaphore reduction operation.
+    /// </summary>
+    enum LaunchDmaReductionOp
+    {
+        RedAdd = 0,
+        RedMin = 1,
+        RedMax = 2,
+        RedInc = 3,
+        RedDec = 4,
+        RedAnd = 5,
+        RedOr = 6,
+        RedXor = 7,
+    }
+
+    /// <summary>
+    /// DMA semaphore reduction format.
+    /// </summary>
+    enum LaunchDmaReductionFormat
+    {
+        Unsigned32 = 0,
+        Signed32 = 1,
+    }
+
+    /// <summary>
+    /// Inline-to-Memory class state.
+    /// </summary>
+    unsafe struct InlineToMemoryClassState
+    {
+#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 fixed uint Reserved114[7];
+        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[9];
+        public uint SetI2mSemaphoreA;
+        public int SetI2mSemaphoreAOffsetUpper => (int)((SetI2mSemaphoreA >> 0) & 0xFF);
+        public uint SetI2mSemaphoreB;
+        public uint SetI2mSemaphoreC;
+        public fixed uint Reserved1E8[2];
+        public uint SetI2mSpareNoop00;
+        public uint SetI2mSpareNoop01;
+        public uint SetI2mSpareNoop02;
+        public uint SetI2mSpareNoop03;
+        public fixed uint Reserved200[3200];
+        public MmeShadowScratch SetMmeShadowScratch;
+#pragma warning restore CS0649
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/MME/AluOperation.cs b/src/Ryujinx.Graphics.Gpu/Engine/MME/AluOperation.cs
new file mode 100644
index 00000000..eeef9c67
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/MME/AluOperation.cs
@@ -0,0 +1,15 @@
+namespace Ryujinx.Graphics.Gpu.Engine.MME
+{
+    /// <summary>
+    /// GPU Macro Arithmetic and Logic unit operation.
+    /// </summary>
+    enum AluOperation
+    {
+        AluReg = 0,
+        AddImmediate = 1,
+        BitfieldReplace = 2,
+        BitfieldExtractLslImm = 3,
+        BitfieldExtractLslReg = 4,
+        ReadImmediate = 5
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/MME/AluRegOperation.cs b/src/Ryujinx.Graphics.Gpu/Engine/MME/AluRegOperation.cs
new file mode 100644
index 00000000..f3e05d38
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/MME/AluRegOperation.cs
@@ -0,0 +1,18 @@
+namespace Ryujinx.Graphics.Gpu.Engine.MME
+{
+    /// <summary>
+    /// GPU Macro Arithmetic and Logic unit binary register-to-register operation.
+    /// </summary>
+    enum AluRegOperation
+    {
+        Add = 0,
+        AddWithCarry = 1,
+        Subtract = 2,
+        SubtractWithBorrow = 3,
+        BitwiseExclusiveOr = 8,
+        BitwiseOr = 9,
+        BitwiseAnd = 10,
+        BitwiseAndNot = 11,
+        BitwiseNotAnd = 12
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/MME/AssignmentOperation.cs b/src/Ryujinx.Graphics.Gpu/Engine/MME/AssignmentOperation.cs
new file mode 100644
index 00000000..dc336026
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/MME/AssignmentOperation.cs
@@ -0,0 +1,17 @@
+namespace Ryujinx.Graphics.Gpu.Engine.MME
+{
+    /// <summary>
+    /// GPU Macro assignment operation.
+    /// </summary>
+    enum AssignmentOperation
+    {
+        IgnoreAndFetch = 0,
+        Move = 1,
+        MoveAndSetMaddr = 2,
+        FetchAndSend = 3,
+        MoveAndSend = 4,
+        FetchAndSetMaddr = 5,
+        MoveAndSetMaddrThenFetchAndSend = 6,
+        MoveAndSetMaddrThenSendHigh = 7
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/MME/IMacroEE.cs b/src/Ryujinx.Graphics.Gpu/Engine/MME/IMacroEE.cs
new file mode 100644
index 00000000..117961db
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/MME/IMacroEE.cs
@@ -0,0 +1,52 @@
+using Ryujinx.Graphics.Device;
+using System;
+using System.Collections.Generic;
+
+namespace Ryujinx.Graphics.Gpu.Engine.MME
+{
+    /// <summary>
+    /// FIFO word.
+    /// </summary>
+    readonly struct FifoWord
+    {
+        /// <summary>
+        /// GPU virtual address where the word is located in memory.
+        /// </summary>
+        public ulong GpuVa { get; }
+
+        /// <summary>
+        /// Word value.
+        /// </summary>
+        public int Word { get; }
+
+        /// <summary>
+        /// Creates a new FIFO word.
+        /// </summary>
+        /// <param name="gpuVa">GPU virtual address where the word is located in memory</param>
+        /// <param name="word">Word value</param>
+        public FifoWord(ulong gpuVa, int word)
+        {
+            GpuVa = gpuVa;
+            Word = word;
+        }
+    }
+
+    /// <summary>
+    /// Macro Execution Engine interface.
+    /// </summary>
+    interface IMacroEE
+    {
+        /// <summary>
+        /// Arguments FIFO.
+        /// </summary>
+        Queue<FifoWord> Fifo { get; }
+
+        /// <summary>
+        /// Should execute the GPU Macro code being passed.
+        /// </summary>
+        /// <param name="code">Code to be executed</param>
+        /// <param name="state">GPU state at the time of the call</param>
+        /// <param name="arg0">First argument to be passed to the GPU Macro</param>
+        void Execute(ReadOnlySpan<int> code, IDeviceState state, int arg0);
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/MME/Macro.cs b/src/Ryujinx.Graphics.Gpu/Engine/MME/Macro.cs
new file mode 100644
index 00000000..12a3ac02
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/MME/Macro.cs
@@ -0,0 +1,101 @@
+using Ryujinx.Graphics.Device;
+using Ryujinx.Graphics.Gpu.Engine.GPFifo;
+using System;
+
+namespace Ryujinx.Graphics.Gpu.Engine.MME
+{
+    /// <summary>
+    /// GPU macro program.
+    /// </summary>
+    struct Macro
+    {
+        /// <summary>
+        /// Word offset of the code on the code memory.
+        /// </summary>
+        public int Position { get; }
+
+        private IMacroEE _executionEngine;
+        private bool _executionPending;
+        private int _argument;
+        private MacroHLEFunctionName _hleFunction;
+
+        /// <summary>
+        /// Creates a new instance of the GPU cached macro program.
+        /// </summary>
+        /// <param name="position">Macro code start position</param>
+        public Macro(int position)
+        {
+            Position = position;
+
+            _executionEngine = null;
+            _executionPending = false;
+            _argument = 0;
+            _hleFunction = MacroHLEFunctionName.None;
+        }
+
+        /// <summary>
+        /// Sets the first argument for the macro call.
+        /// </summary>
+        /// <param name="context">GPU context where the macro code is being executed</param>
+        /// <param name="processor">GPU GP FIFO command processor</param>
+        /// <param name="code">Code to be executed</param>
+        /// <param name="argument">First argument</param>
+        public void StartExecution(GpuContext context, GPFifoProcessor processor, ReadOnlySpan<int> code, int argument)
+        {
+            _argument = argument;
+
+            _executionPending = true;
+
+            if (_executionEngine == null)
+            {
+                if (GraphicsConfig.EnableMacroHLE && MacroHLETable.TryGetMacroHLEFunction(code.Slice(Position), context.Capabilities, out _hleFunction))
+                {
+                    _executionEngine = new MacroHLE(processor, _hleFunction);
+                }
+                else if (GraphicsConfig.EnableMacroJit)
+                {
+                    _executionEngine = new MacroJit();
+                }
+                else
+                {
+                    _executionEngine = new MacroInterpreter();
+                }
+            }
+
+            // We don't consume the parameter buffer value, so we don't need to flush it.
+            // Doing so improves performance if the value was written by a GPU shader.
+            if (_hleFunction == MacroHLEFunctionName.DrawElementsIndirect)
+            {
+                context.GPFifo.SetFlushSkips(1);
+            }
+            else if (_hleFunction == MacroHLEFunctionName.MultiDrawElementsIndirectCount)
+            {
+                context.GPFifo.SetFlushSkips(2);
+            }
+        }
+
+        /// <summary>
+        /// Starts executing the macro program code.
+        /// </summary>
+        /// <param name="code">Program code</param>
+        /// <param name="state">Current GPU state</param>
+        public void Execute(ReadOnlySpan<int> code, IDeviceState state)
+        {
+            if (_executionPending)
+            {
+                _executionPending = false;
+                _executionEngine?.Execute(code.Slice(Position), state, _argument);
+            }
+        }
+
+        /// <summary>
+        /// Pushes an argument to the macro call argument FIFO.
+        /// </summary>
+        /// <param name="gpuVa">GPU virtual address where the command word is located</param>
+        /// <param name="argument">Argument to be pushed</param>
+        public void PushArgument(ulong gpuVa, int argument)
+        {
+            _executionEngine?.Fifo.Enqueue(new FifoWord(gpuVa, argument));
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroHLE.cs b/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroHLE.cs
new file mode 100644
index 00000000..8630bbc4
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroHLE.cs
@@ -0,0 +1,341 @@
+using Ryujinx.Common.Logging;
+using Ryujinx.Graphics.Device;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.GPFifo;
+using System;
+using System.Collections.Generic;
+
+namespace Ryujinx.Graphics.Gpu.Engine.MME
+{
+    /// <summary>
+    /// Macro High-level emulation.
+    /// </summary>
+    class MacroHLE : IMacroEE
+    {
+        private const int ColorLayerCountOffset = 0x818;
+        private const int ColorStructSize = 0x40;
+        private const int ZetaLayerCountOffset = 0x1230;
+
+        private const int IndirectDataEntrySize = 0x10;
+        private const int IndirectIndexedDataEntrySize = 0x14;
+
+        private readonly GPFifoProcessor _processor;
+        private readonly MacroHLEFunctionName _functionName;
+
+        /// <summary>
+        /// Arguments FIFO.
+        /// </summary>
+        public Queue<FifoWord> Fifo { get; }
+
+        /// <summary>
+        /// Creates a new instance of the HLE macro handler.
+        /// </summary>
+        /// <param name="processor">GPU GP FIFO command processor</param>
+        /// <param name="functionName">Name of the HLE macro function to be called</param>
+        public MacroHLE(GPFifoProcessor processor, MacroHLEFunctionName functionName)
+        {
+            _processor = processor;
+            _functionName = functionName;
+
+            Fifo = new Queue<FifoWord>();
+        }
+
+        /// <summary>
+        /// Executes a macro program until it exits.
+        /// </summary>
+        /// <param name="code">Code of the program to execute</param>
+        /// <param name="state">GPU state at the time of the call</param>
+        /// <param name="arg0">Optional argument passed to the program, 0 if not used</param>
+        public void Execute(ReadOnlySpan<int> code, IDeviceState state, int arg0)
+        {
+            switch (_functionName)
+            {
+                case MacroHLEFunctionName.ClearColor:
+                    ClearColor(state, arg0);
+                    break;
+                case MacroHLEFunctionName.ClearDepthStencil:
+                    ClearDepthStencil(state, arg0);
+                    break;
+                case MacroHLEFunctionName.DrawArraysInstanced:
+                    DrawArraysInstanced(state, arg0);
+                    break;
+                case MacroHLEFunctionName.DrawElementsInstanced:
+                    DrawElementsInstanced(state, arg0);
+                    break;
+                case MacroHLEFunctionName.DrawElementsIndirect:
+                    DrawElementsIndirect(state, arg0);
+                    break;
+                case MacroHLEFunctionName.MultiDrawElementsIndirectCount:
+                    MultiDrawElementsIndirectCount(state, arg0);
+                    break;
+                default:
+                    throw new NotImplementedException(_functionName.ToString());
+            }
+
+            // It should be empty at this point, but clear it just to be safe.
+            Fifo.Clear();
+        }
+
+        /// <summary>
+        /// Clears one bound color target.
+        /// </summary>
+        /// <param name="state">GPU state at the time of the call</param>
+        /// <param name="arg0">First argument of the call</param>
+        private void ClearColor(IDeviceState state, int arg0)
+        {
+            int index = (arg0 >> 6) & 0xf;
+            int layerCount = state.Read(ColorLayerCountOffset + index * ColorStructSize);
+
+            _processor.ThreedClass.Clear(arg0, layerCount);
+        }
+
+        /// <summary>
+        /// Clears the current depth-stencil target.
+        /// </summary>
+        /// <param name="state">GPU state at the time of the call</param>
+        /// <param name="arg0">First argument of the call</param>
+        private void ClearDepthStencil(IDeviceState state, int arg0)
+        {
+            int layerCount = state.Read(ZetaLayerCountOffset);
+
+            _processor.ThreedClass.Clear(arg0, layerCount);
+        }
+
+        /// <summary>
+        /// Performs a draw.
+        /// </summary>
+        /// <param name="state">GPU state at the time of the call</param>
+        /// <param name="arg0">First argument of the call</param>
+        private void DrawArraysInstanced(IDeviceState state, int arg0)
+        {
+            var topology = (PrimitiveTopology)arg0;
+
+            var count = FetchParam();
+            var instanceCount = FetchParam();
+            var firstVertex = FetchParam();
+            var firstInstance = FetchParam();
+
+            if (ShouldSkipDraw(state, instanceCount.Word))
+            {
+                return;
+            }
+
+            _processor.ThreedClass.Draw(
+                topology,
+                count.Word,
+                instanceCount.Word,
+                0,
+                firstVertex.Word,
+                firstInstance.Word,
+                indexed: false);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw.
+        /// </summary>
+        /// <param name="state">GPU state at the time of the call</param>
+        /// <param name="arg0">First argument of the call</param>
+        private void DrawElementsInstanced(IDeviceState state, int arg0)
+        {
+            var topology = (PrimitiveTopology)arg0;
+
+            var count = FetchParam();
+            var instanceCount = FetchParam();
+            var firstIndex = FetchParam();
+            var firstVertex = FetchParam();
+            var firstInstance = FetchParam();
+
+            if (ShouldSkipDraw(state, instanceCount.Word))
+            {
+                return;
+            }
+
+            _processor.ThreedClass.Draw(
+                topology,
+                count.Word,
+                instanceCount.Word,
+                firstIndex.Word,
+                firstVertex.Word,
+                firstInstance.Word,
+                indexed: true);
+        }
+
+        /// <summary>
+        /// Performs a indirect indexed draw, with parameters from a GPU buffer.
+        /// </summary>
+        /// <param name="state">GPU state at the time of the call</param>
+        /// <param name="arg0">First argument of the call</param>
+        private void DrawElementsIndirect(IDeviceState state, int arg0)
+        {
+            var topology = (PrimitiveTopology)arg0;
+
+            var count = FetchParam();
+            var instanceCount = FetchParam();
+            var firstIndex = FetchParam();
+            var firstVertex = FetchParam();
+            var firstInstance = FetchParam();
+
+            ulong indirectBufferGpuVa = count.GpuVa;
+
+            var bufferCache = _processor.MemoryManager.Physical.BufferCache;
+
+            bool useBuffer = bufferCache.CheckModified(_processor.MemoryManager, indirectBufferGpuVa, IndirectIndexedDataEntrySize, out ulong indirectBufferAddress);
+
+            if (useBuffer)
+            {
+                int indexCount = firstIndex.Word + count.Word;
+
+                _processor.ThreedClass.DrawIndirect(
+                    topology,
+                    indirectBufferAddress,
+                    0,
+                    1,
+                    IndirectIndexedDataEntrySize,
+                    indexCount,
+                    Threed.IndirectDrawType.DrawIndexedIndirect);
+            }
+            else
+            {
+                if (ShouldSkipDraw(state, instanceCount.Word))
+                {
+                    return;
+                }
+
+                _processor.ThreedClass.Draw(
+                    topology,
+                    count.Word,
+                    instanceCount.Word,
+                    firstIndex.Word,
+                    firstVertex.Word,
+                    firstInstance.Word,
+                    indexed: true);
+            }
+        }
+
+        /// <summary>
+        /// Performs a indirect indexed multi-draw, with parameters from a GPU buffer.
+        /// </summary>
+        /// <param name="state">GPU state at the time of the call</param>
+        /// <param name="arg0">First argument of the call</param>
+        private void MultiDrawElementsIndirectCount(IDeviceState state, int arg0)
+        {
+            int arg1 = FetchParam().Word;
+            int arg2 = FetchParam().Word;
+            int arg3 = FetchParam().Word;
+
+            int startDraw = arg0;
+            int endDraw = arg1;
+            var topology = (PrimitiveTopology)arg2;
+            int paddingWords = arg3;
+            int stride = paddingWords * 4 + 0x14;
+
+            ulong parameterBufferGpuVa = FetchParam().GpuVa;
+
+            int maxDrawCount = endDraw - startDraw;
+
+            if (startDraw != 0)
+            {
+                int drawCount = _processor.MemoryManager.Read<int>(parameterBufferGpuVa, tracked: true);
+
+                // Calculate maximum draw count based on the previous draw count and current draw count.
+                if ((uint)drawCount <= (uint)startDraw)
+                {
+                    // The start draw is past our total draw count, so all draws were already performed.
+                    maxDrawCount = 0;
+                }
+                else
+                {
+                    // Perform just the missing number of draws.
+                    maxDrawCount = (int)Math.Min((uint)maxDrawCount, (uint)(drawCount - startDraw));
+                }
+            }
+
+            if (maxDrawCount == 0)
+            {
+                Fifo.Clear();
+                return;
+            }
+
+            ulong indirectBufferGpuVa = 0;
+            int indexCount = 0;
+
+            for (int i = 0; i < maxDrawCount; i++)
+            {
+                var count = FetchParam();
+                var instanceCount = FetchParam();
+                var firstIndex = FetchParam();
+                var firstVertex = FetchParam();
+                var firstInstance = FetchParam();
+
+                if (i == 0)
+                {
+                    indirectBufferGpuVa = count.GpuVa;
+                }
+
+                indexCount = Math.Max(indexCount, count.Word + firstIndex.Word);
+
+                if (i != maxDrawCount - 1)
+                {
+                    for (int j = 0; j < paddingWords; j++)
+                    {
+                        FetchParam();
+                    }
+                }
+            }
+
+            var bufferCache = _processor.MemoryManager.Physical.BufferCache;
+
+            ulong indirectBufferSize = (ulong)maxDrawCount * (ulong)stride;
+
+            ulong indirectBufferAddress = bufferCache.TranslateAndCreateBuffer(_processor.MemoryManager, indirectBufferGpuVa, indirectBufferSize);
+            ulong parameterBufferAddress = bufferCache.TranslateAndCreateBuffer(_processor.MemoryManager, parameterBufferGpuVa, 4);
+
+            _processor.ThreedClass.DrawIndirect(
+                topology,
+                indirectBufferAddress,
+                parameterBufferAddress,
+                maxDrawCount,
+                stride,
+                indexCount,
+                Threed.IndirectDrawType.DrawIndexedIndirectCount);
+        }
+
+        /// <summary>
+        /// Checks if the draw should be skipped, because the masked instance count is zero.
+        /// </summary>
+        /// <param name="state">Current GPU state</param>
+        /// <param name="instanceCount">Draw instance count</param>
+        /// <returns>True if the draw should be skipped, false otherwise</returns>
+        private static bool ShouldSkipDraw(IDeviceState state, int instanceCount)
+        {
+            return (Read(state, 0xd1b) & instanceCount) == 0;
+        }
+
+        /// <summary>
+        /// Fetches a arguments from the arguments FIFO.
+        /// </summary>
+        /// <returns>The call argument, or a 0 value with null address if the FIFO is empty</returns>
+        private FifoWord FetchParam()
+        {
+            if (!Fifo.TryDequeue(out var value))
+            {
+                Logger.Warning?.Print(LogClass.Gpu, "Macro attempted to fetch an inexistent argument.");
+
+                return new FifoWord(0UL, 0);
+            }
+
+            return value;
+        }
+
+        /// <summary>
+        /// Reads data from a GPU register.
+        /// </summary>
+        /// <param name="state">Current GPU state</param>
+        /// <param name="reg">Register offset to read</param>
+        /// <returns>GPU register value</returns>
+        private static int Read(IDeviceState state, int reg)
+        {
+            return state.Read(reg * 4);
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroHLEFunctionName.cs b/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroHLEFunctionName.cs
new file mode 100644
index 00000000..751867fc
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroHLEFunctionName.cs
@@ -0,0 +1,16 @@
+namespace Ryujinx.Graphics.Gpu.Engine.MME
+{
+    /// <summary>
+    /// Name of the High-level implementation of a Macro function.
+    /// </summary>
+    enum MacroHLEFunctionName
+    {
+        None,
+        ClearColor,
+        ClearDepthStencil,
+        DrawArraysInstanced,
+        DrawElementsInstanced,
+        DrawElementsIndirect,
+        MultiDrawElementsIndirectCount
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroHLETable.cs b/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroHLETable.cs
new file mode 100644
index 00000000..719e170f
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroHLETable.cs
@@ -0,0 +1,113 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.GAL;
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.MME
+{
+    /// <summary>
+    /// Table with information about High-level implementations of GPU Macro code.
+    /// </summary>
+    static class MacroHLETable
+    {
+        /// <summary>
+        /// Macroo High-level implementation table entry.
+        /// </summary>
+        readonly struct TableEntry
+        {
+            /// <summary>
+            /// Name of the Macro function.
+            /// </summary>
+            public MacroHLEFunctionName Name { get; }
+
+            /// <summary>
+            /// Hash of the original binary Macro function code.
+            /// </summary>
+            public Hash128 Hash { get; }
+
+            /// <summary>
+            /// Size (in bytes) of the original binary Macro function code.
+            /// </summary>
+            public int Length { get; }
+
+            /// <summary>
+            /// Creates a new table entry.
+            /// </summary>
+            /// <param name="name">Name of the Macro function</param>
+            /// <param name="hash">Hash of the original binary Macro function code</param>
+            /// <param name="length">Size (in bytes) of the original binary Macro function code</param>
+            public TableEntry(MacroHLEFunctionName name, Hash128 hash, int length)
+            {
+                Name = name;
+                Hash = hash;
+                Length = length;
+            }
+        }
+
+        private static readonly TableEntry[] _table = new TableEntry[]
+        {
+            new TableEntry(MacroHLEFunctionName.ClearColor, new Hash128(0xA9FB28D1DC43645A, 0xB177E5D2EAE67FB0), 0x28),
+            new TableEntry(MacroHLEFunctionName.ClearDepthStencil, new Hash128(0x1B96CB77D4879F4F, 0x8557032FE0C965FB), 0x24),
+            new TableEntry(MacroHLEFunctionName.DrawArraysInstanced, new Hash128(0x197FB416269DBC26, 0x34288C01DDA82202), 0x48),
+            new TableEntry(MacroHLEFunctionName.DrawElementsInstanced, new Hash128(0x1A501FD3D54EC8E0, 0x6CF570CF79DA74D6), 0x5c),
+            new TableEntry(MacroHLEFunctionName.DrawElementsIndirect, new Hash128(0x86A3E8E903AF8F45, 0xD35BBA07C23860A4), 0x7c),
+            new TableEntry(MacroHLEFunctionName.MultiDrawElementsIndirectCount, new Hash128(0x890AF57ED3FB1C37, 0x35D0C95C61F5386F), 0x19C)
+        };
+
+        /// <summary>
+        /// Checks if the host supports all features required by the HLE macro.
+        /// </summary>
+        /// <param name="caps">Host capabilities</param>
+        /// <param name="name">Name of the HLE macro to be checked</param>
+        /// <returns>True if the host supports the HLE macro, false otherwise</returns>
+        private static bool IsMacroHLESupported(Capabilities caps, MacroHLEFunctionName name)
+        {
+            if (name == MacroHLEFunctionName.ClearColor ||
+                name == MacroHLEFunctionName.ClearDepthStencil ||
+                name == MacroHLEFunctionName.DrawArraysInstanced ||
+                name == MacroHLEFunctionName.DrawElementsInstanced ||
+                name == MacroHLEFunctionName.DrawElementsIndirect)
+            {
+                return true;
+            }
+            else if (name == MacroHLEFunctionName.MultiDrawElementsIndirectCount)
+            {
+                return caps.SupportsIndirectParameters;
+            }
+
+            return false;
+        }
+
+        /// <summary>
+        /// Checks if there's a fast, High-level implementation of the specified Macro code available.
+        /// </summary>
+        /// <param name="code">Macro code to be checked</param>
+        /// <param name="caps">Renderer capabilities to check for this macro HLE support</param>
+        /// <param name="name">Name of the function if a implementation is available and supported, otherwise <see cref="MacroHLEFunctionName.None"/></param>
+        /// <returns>True if there is a implementation available and supported, false otherwise</returns>
+        public static bool TryGetMacroHLEFunction(ReadOnlySpan<int> code, Capabilities caps, out MacroHLEFunctionName name)
+        {
+            var mc = MemoryMarshal.Cast<int, byte>(code);
+
+            for (int i = 0; i < _table.Length; i++)
+            {
+                ref var entry = ref _table[i];
+
+                var hash = XXHash128.ComputeHash(mc.Slice(0, entry.Length));
+                if (hash == entry.Hash)
+                {
+                    if (IsMacroHLESupported(caps, entry.Name))
+                    {
+                        name = entry.Name;
+                        return true;
+                    }
+
+                    break;
+                }
+            }
+
+            name = MacroHLEFunctionName.None;
+            return false;
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroInterpreter.cs b/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroInterpreter.cs
new file mode 100644
index 00000000..df6ee040
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroInterpreter.cs
@@ -0,0 +1,400 @@
+using Ryujinx.Common.Logging;
+using Ryujinx.Graphics.Device;
+using System;
+using System.Collections.Generic;
+
+namespace Ryujinx.Graphics.Gpu.Engine.MME
+{
+    /// <summary>
+    /// Macro code interpreter.
+    /// </summary>
+    class MacroInterpreter : IMacroEE
+    {
+        /// <summary>
+        /// Arguments FIFO.
+        /// </summary>
+        public Queue<FifoWord> Fifo { get; }
+
+        private int[] _gprs;
+
+        private int _methAddr;
+        private int _methIncr;
+
+        private bool _carry;
+
+        private int _opCode;
+        private int _pipeOp;
+
+        private bool _ignoreExitFlag;
+
+        private int _pc;
+
+        /// <summary>
+        /// Creates a new instance of the macro code interpreter.
+        /// </summary>
+        public MacroInterpreter()
+        {
+            Fifo = new Queue<FifoWord>();
+
+            _gprs = new int[8];
+        }
+
+        /// <summary>
+        /// Executes a macro program until it exits.
+        /// </summary>
+        /// <param name="code">Code of the program to execute</param>
+        /// <param name="state">Current GPU state</param>
+        /// <param name="arg0">Optional argument passed to the program, 0 if not used</param>
+        public void Execute(ReadOnlySpan<int> code, IDeviceState state, int arg0)
+        {
+            Reset();
+
+            _gprs[1] = arg0;
+
+            _pc = 0;
+
+            FetchOpCode(code);
+
+            while (Step(code, state))
+            {
+            }
+
+            // Due to the delay slot, we still need to execute
+            // one more instruction before we actually exit.
+            Step(code, state);
+        }
+
+        /// <summary>
+        /// Resets the internal interpreter state.
+        /// Call each time you run a new program.
+        /// </summary>
+        private void Reset()
+        {
+            for (int index = 0; index < _gprs.Length; index++)
+            {
+                _gprs[index] = 0;
+            }
+
+            _methAddr = 0;
+            _methIncr = 0;
+
+            _carry = false;
+        }
+
+        /// <summary>
+        /// Executes a single instruction of the program.
+        /// </summary>
+        /// <param name="code">Program code to execute</param>
+        /// <param name="state">Current GPU state</param>
+        /// <returns>True to continue execution, false if the program exited</returns>
+        private bool Step(ReadOnlySpan<int> code, IDeviceState state)
+        {
+            int baseAddr = _pc - 1;
+
+            FetchOpCode(code);
+
+            if ((_opCode & 7) < 7)
+            {
+                // Operation produces a value.
+                AssignmentOperation asgOp = (AssignmentOperation)((_opCode >> 4) & 7);
+
+                int result = GetAluResult(state);
+
+                switch (asgOp)
+                {
+                    // Fetch parameter and ignore result.
+                    case AssignmentOperation.IgnoreAndFetch:
+                        SetDstGpr(FetchParam());
+                        break;
+                    // Move result.
+                    case AssignmentOperation.Move:
+                        SetDstGpr(result);
+                        break;
+                    // Move result and use as Method Address.
+                    case AssignmentOperation.MoveAndSetMaddr:
+                        SetDstGpr(result);
+                        SetMethAddr(result);
+                        break;
+                    // Fetch parameter and send result.
+                    case AssignmentOperation.FetchAndSend:
+                        SetDstGpr(FetchParam());
+                        Send(state, result);
+                        break;
+                    // Move and send result.
+                    case AssignmentOperation.MoveAndSend:
+                        SetDstGpr(result);
+                        Send(state, result);
+                        break;
+                    // Fetch parameter and use result as Method Address.
+                    case AssignmentOperation.FetchAndSetMaddr:
+                        SetDstGpr(FetchParam());
+                        SetMethAddr(result);
+                        break;
+                    // Move result and use as Method Address, then fetch and send parameter.
+                    case AssignmentOperation.MoveAndSetMaddrThenFetchAndSend:
+                        SetDstGpr(result);
+                        SetMethAddr(result);
+                        Send(state, FetchParam());
+                        break;
+                    // Move result and use as Method Address, then send bits 17:12 of result.
+                    case AssignmentOperation.MoveAndSetMaddrThenSendHigh:
+                        SetDstGpr(result);
+                        SetMethAddr(result);
+                        Send(state, (result >> 12) & 0x3f);
+                        break;
+                }
+            }
+            else
+            {
+                // Branch.
+                bool onNotZero = ((_opCode >> 4) & 1) != 0;
+
+                bool taken = onNotZero
+                    ? GetGprA() != 0
+                    : GetGprA() == 0;
+
+                if (taken)
+                {
+                    _pc = baseAddr + GetImm();
+
+                    bool noDelays = (_opCode & 0x20) != 0;
+
+                    if (noDelays)
+                    {
+                        FetchOpCode(code);
+                    }
+                    else
+                    {
+                        // The delay slot instruction exit flag should be ignored.
+                        _ignoreExitFlag = true;
+                    }
+
+                    return true;
+                }
+            }
+
+            bool exit = (_opCode & 0x80) != 0 && !_ignoreExitFlag;
+
+            _ignoreExitFlag = false;
+
+            return !exit;
+        }
+
+        /// <summary>
+        /// Fetches a single operation code from the program code.
+        /// </summary>
+        /// <param name="code">Program code</param>
+        private void FetchOpCode(ReadOnlySpan<int> code)
+        {
+            _opCode = _pipeOp;
+            _pipeOp = code[_pc++];
+        }
+
+        /// <summary>
+        /// Gets the result of the current Arithmetic and Logic unit operation.
+        /// </summary>
+        /// <param name="state">Current GPU state</param>
+        /// <returns>Operation result</returns>
+        private int GetAluResult(IDeviceState state)
+        {
+            AluOperation op = (AluOperation)(_opCode & 7);
+
+            switch (op)
+            {
+                case AluOperation.AluReg:
+                    return GetAluResult((AluRegOperation)((_opCode >> 17) & 0x1f), GetGprA(), GetGprB());
+
+                case AluOperation.AddImmediate:
+                    return GetGprA() + GetImm();
+
+                case AluOperation.BitfieldReplace:
+                case AluOperation.BitfieldExtractLslImm:
+                case AluOperation.BitfieldExtractLslReg:
+                    int bfSrcBit = (_opCode >> 17) & 0x1f;
+                    int bfSize = (_opCode >> 22) & 0x1f;
+                    int bfDstBit = (_opCode >> 27) & 0x1f;
+
+                    int bfMask = (1 << bfSize) - 1;
+
+                    int dst = GetGprA();
+                    int src = GetGprB();
+
+                    switch (op)
+                    {
+                        case AluOperation.BitfieldReplace:
+                            src = (int)((uint)src >> bfSrcBit) & bfMask;
+
+                            dst &= ~(bfMask << bfDstBit);
+
+                            dst |= src << bfDstBit;
+
+                            return dst;
+
+                        case AluOperation.BitfieldExtractLslImm:
+                            src = (int)((uint)src >> dst) & bfMask;
+
+                            return src << bfDstBit;
+
+                        case AluOperation.BitfieldExtractLslReg:
+                            src = (int)((uint)src >> bfSrcBit) & bfMask;
+
+                            return src << dst;
+                    }
+
+                    break;
+
+                case AluOperation.ReadImmediate:
+                    return Read(state, GetGprA() + GetImm());
+            }
+
+            throw new InvalidOperationException($"Invalid operation \"{op}\" on instruction 0x{_opCode:X8}.");
+        }
+
+        /// <summary>
+        /// Gets the result of an Arithmetic and Logic operation using registers.
+        /// </summary>
+        /// <param name="aluOp">Arithmetic and Logic unit operation with registers</param>
+        /// <param name="a">First operand value</param>
+        /// <param name="b">Second operand value</param>
+        /// <returns>Operation result</returns>
+        private int GetAluResult(AluRegOperation aluOp, int a, int b)
+        {
+            ulong result;
+
+            switch (aluOp)
+            {
+                case AluRegOperation.Add:
+                    result = (ulong)a + (ulong)b;
+
+                    _carry = result > 0xffffffff;
+
+                    return (int)result;
+
+                case AluRegOperation.AddWithCarry:
+                    result = (ulong)a + (ulong)b + (_carry ? 1UL : 0UL);
+
+                    _carry = result > 0xffffffff;
+
+                    return (int)result;
+
+                case AluRegOperation.Subtract:
+                    result = (ulong)a - (ulong)b;
+
+                    _carry = result < 0x100000000;
+
+                    return (int)result;
+
+                case AluRegOperation.SubtractWithBorrow:
+                    result = (ulong)a - (ulong)b - (_carry ? 0UL : 1UL);
+
+                    _carry = result < 0x100000000;
+
+                    return (int)result;
+
+                case AluRegOperation.BitwiseExclusiveOr: return a ^ b;
+                case AluRegOperation.BitwiseOr: return a | b;
+                case AluRegOperation.BitwiseAnd: return a & b;
+                case AluRegOperation.BitwiseAndNot: return a & ~b;
+                case AluRegOperation.BitwiseNotAnd: return ~(a & b);
+            }
+
+            throw new InvalidOperationException($"Invalid operation \"{aluOp}\" on instruction 0x{_opCode:X8}.");
+        }
+
+        /// <summary>
+        /// Extracts a 32-bits signed integer constant from the current operation code.
+        /// </summary>
+        /// <returns>The 32-bits immediate value encoded at the current operation code</returns>
+        private int GetImm()
+        {
+            // Note: The immediate is signed, the sign-extension is intended here.
+            return _opCode >> 14;
+        }
+
+        /// <summary>
+        /// Sets the current method address, for method calls.
+        /// </summary>
+        /// <param name="value">Packed address and increment value</param>
+        private void SetMethAddr(int value)
+        {
+            _methAddr = (value >> 0) & 0xfff;
+            _methIncr = (value >> 12) & 0x3f;
+        }
+
+        /// <summary>
+        /// Sets the destination register value.
+        /// </summary>
+        /// <param name="value">Value to set (usually the operation result)</param>
+        private void SetDstGpr(int value)
+        {
+            _gprs[(_opCode >> 8) & 7] = value;
+        }
+
+        /// <summary>
+        /// Gets first operand value from the respective register.
+        /// </summary>
+        /// <returns>Operand value</returns>
+        private int GetGprA()
+        {
+            return GetGprValue((_opCode >> 11) & 7);
+        }
+
+        /// <summary>
+        /// Gets second operand value from the respective register.
+        /// </summary>
+        /// <returns>Operand value</returns>
+        private int GetGprB()
+        {
+            return GetGprValue((_opCode >> 14) & 7);
+        }
+
+        /// <summary>
+        /// Gets the value from a register, or 0 if the R0 register is specified.
+        /// </summary>
+        /// <param name="index">Index of the register</param>
+        /// <returns>Register value</returns>
+        private int GetGprValue(int index)
+        {
+            return index != 0 ? _gprs[index] : 0;
+        }
+
+        /// <summary>
+        /// Fetches a call argument from the call argument FIFO.
+        /// </summary>
+        /// <returns>The call argument, or 0 if the FIFO is empty</returns>
+        private int FetchParam()
+        {
+            if (!Fifo.TryDequeue(out var value))
+            {
+                Logger.Warning?.Print(LogClass.Gpu, "Macro attempted to fetch an inexistent argument.");
+
+                return 0;
+            }
+
+            return value.Word;
+        }
+
+        /// <summary>
+        /// Reads data from a GPU register.
+        /// </summary>
+        /// <param name="state">Current GPU state</param>
+        /// <param name="reg">Register offset to read</param>
+        /// <returns>GPU register value</returns>
+        private int Read(IDeviceState state, int reg)
+        {
+            return state.Read(reg * 4);
+        }
+
+        /// <summary>
+        /// Performs a GPU method call.
+        /// </summary>
+        /// <param name="state">Current GPU state</param>
+        /// <param name="value">Call argument</param>
+        private void Send(IDeviceState state, int value)
+        {
+            state.Write(_methAddr * 4, value);
+
+            _methAddr += _methIncr;
+        }
+    }
+}
\ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroJit.cs b/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroJit.cs
new file mode 100644
index 00000000..4077f74e
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroJit.cs
@@ -0,0 +1,39 @@
+using Ryujinx.Graphics.Device;
+using System;
+using System.Collections.Generic;
+
+namespace Ryujinx.Graphics.Gpu.Engine.MME
+{
+    /// <summary>
+    /// Represents a execution engine that uses a Just-in-Time compiler for fast execution.
+    /// </summary>
+    class MacroJit : IMacroEE
+    {
+        private readonly MacroJitContext _context = new MacroJitContext();
+
+        /// <summary>
+        /// Arguments FIFO.
+        /// </summary>
+        public Queue<FifoWord> Fifo => _context.Fifo;
+
+        private MacroJitCompiler.MacroExecute _execute;
+
+        /// <summary>
+        /// Executes a macro program until it exits.
+        /// </summary>
+        /// <param name="code">Code of the program to execute</param>
+        /// <param name="state">Current GPU state</param>
+        /// <param name="arg0">Optional argument passed to the program, 0 if not used</param>
+        public void Execute(ReadOnlySpan<int> code, IDeviceState state, int arg0)
+        {
+            if (_execute == null)
+            {
+                MacroJitCompiler compiler = new MacroJitCompiler();
+
+                _execute = compiler.Compile(code);
+            }
+
+            _execute(_context, state, arg0);
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroJitCompiler.cs b/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroJitCompiler.cs
new file mode 100644
index 00000000..f8d839fa
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroJitCompiler.cs
@@ -0,0 +1,517 @@
+using Ryujinx.Graphics.Device;
+using System;
+using System.Collections.Generic;
+using System.Reflection.Emit;
+
+namespace Ryujinx.Graphics.Gpu.Engine.MME
+{
+    /// <summary>
+    /// Represents a Macro Just-in-Time compiler.
+    /// </summary>R
+    class MacroJitCompiler
+    {
+        private readonly DynamicMethod _meth;
+        private readonly ILGenerator _ilGen;
+        private readonly LocalBuilder[] _gprs;
+        private readonly LocalBuilder _carry;
+        private readonly LocalBuilder _methAddr;
+        private readonly LocalBuilder _methIncr;
+
+        /// <summary>
+        /// Creates a new instance of the Macro Just-in-Time compiler.
+        /// </summary>
+        public MacroJitCompiler()
+        {
+            _meth = new DynamicMethod("Macro", typeof(void), new Type[] { typeof(MacroJitContext), typeof(IDeviceState), typeof(int) });
+            _ilGen = _meth.GetILGenerator();
+            _gprs = new LocalBuilder[8];
+
+            for (int i = 1; i < 8; i++)
+            {
+                _gprs[i] = _ilGen.DeclareLocal(typeof(int));
+            }
+
+            _carry = _ilGen.DeclareLocal(typeof(int));
+            _methAddr = _ilGen.DeclareLocal(typeof(int));
+            _methIncr = _ilGen.DeclareLocal(typeof(int));
+
+            _ilGen.Emit(OpCodes.Ldarg_2);
+            _ilGen.Emit(OpCodes.Stloc, _gprs[1]);
+        }
+
+        public delegate void MacroExecute(MacroJitContext context, IDeviceState state, int arg0);
+
+        /// <summary>
+        /// Translates a new piece of GPU Macro code into host executable code.
+        /// </summary>
+        /// <param name="code">Code to be translated</param>
+        /// <returns>Delegate of the host compiled code</returns>
+        public MacroExecute Compile(ReadOnlySpan<int> code)
+        {
+            Dictionary<int, Label> labels = new Dictionary<int, Label>();
+
+            int lastTarget = 0;
+            int i;
+
+            // Collect all branch targets.
+            for (i = 0; i < code.Length; i++)
+            {
+                int opCode = code[i];
+
+                if ((opCode & 7) == 7)
+                {
+                    int target = i + (opCode >> 14);
+
+                    if (!labels.ContainsKey(target))
+                    {
+                        labels.Add(target, _ilGen.DefineLabel());
+                    }
+
+                    if (lastTarget < target)
+                    {
+                        lastTarget = target;
+                    }
+                }
+
+                bool exit = (opCode & 0x80) != 0;
+
+                if (exit && i >= lastTarget)
+                {
+                    break;
+                }
+            }
+
+            // Code generation.
+            for (i = 0; i < code.Length; i++)
+            {
+                if (labels.TryGetValue(i, out Label label))
+                {
+                    _ilGen.MarkLabel(label);
+                }
+
+                Emit(code, i, labels);
+
+                int opCode = code[i];
+
+                bool exit = (opCode & 0x80) != 0;
+
+                if (exit)
+                {
+                    Emit(code, i + 1, labels);
+                    _ilGen.Emit(OpCodes.Ret);
+
+                    if (i >= lastTarget)
+                    {
+                        break;
+                    }
+                }
+            }
+
+            if (i == code.Length)
+            {
+                _ilGen.Emit(OpCodes.Ret);
+            }
+
+            return _meth.CreateDelegate<MacroExecute>();
+        }
+
+        /// <summary>
+        /// Emits IL equivalent to the Macro instruction at a given offset.
+        /// </summary>
+        /// <param name="code">GPU Macro code</param>
+        /// <param name="offset">Offset, in words, where the instruction is located</param>
+        /// <param name="labels">Labels for Macro branch targets, used by branch instructions</param>
+        private void Emit(ReadOnlySpan<int> code, int offset, Dictionary<int, Label> labels)
+        {
+            int opCode = code[offset];
+
+            if ((opCode & 7) < 7)
+            {
+                // Operation produces a value.
+                AssignmentOperation asgOp = (AssignmentOperation)((opCode >> 4) & 7);
+
+                EmitAluOp(opCode);
+
+                switch (asgOp)
+                {
+                    // Fetch parameter and ignore result.
+                    case AssignmentOperation.IgnoreAndFetch:
+                        _ilGen.Emit(OpCodes.Pop);
+                        EmitFetchParam();
+                        EmitStoreDstGpr(opCode);
+                        break;
+                    // Move result.
+                    case AssignmentOperation.Move:
+                        EmitStoreDstGpr(opCode);
+                        break;
+                    // Move result and use as Method Address.
+                    case AssignmentOperation.MoveAndSetMaddr:
+                        _ilGen.Emit(OpCodes.Dup);
+                        EmitStoreDstGpr(opCode);
+                        EmitStoreMethAddr();
+                        break;
+                    // Fetch parameter and send result.
+                    case AssignmentOperation.FetchAndSend:
+                        EmitFetchParam();
+                        EmitStoreDstGpr(opCode);
+                        EmitSend();
+                        break;
+                    // Move and send result.
+                    case AssignmentOperation.MoveAndSend:
+                        _ilGen.Emit(OpCodes.Dup);
+                        EmitStoreDstGpr(opCode);
+                        EmitSend();
+                        break;
+                    // Fetch parameter and use result as Method Address.
+                    case AssignmentOperation.FetchAndSetMaddr:
+                        EmitFetchParam();
+                        EmitStoreDstGpr(opCode);
+                        EmitStoreMethAddr();
+                        break;
+                    // Move result and use as Method Address, then fetch and send parameter.
+                    case AssignmentOperation.MoveAndSetMaddrThenFetchAndSend:
+                        _ilGen.Emit(OpCodes.Dup);
+                        EmitStoreDstGpr(opCode);
+                        EmitStoreMethAddr();
+                        EmitFetchParam();
+                        EmitSend();
+                        break;
+                    // Move result and use as Method Address, then send bits 17:12 of result.
+                    case AssignmentOperation.MoveAndSetMaddrThenSendHigh:
+                        _ilGen.Emit(OpCodes.Dup);
+                        _ilGen.Emit(OpCodes.Dup);
+                        EmitStoreDstGpr(opCode);
+                        EmitStoreMethAddr();
+                        _ilGen.Emit(OpCodes.Ldc_I4, 12);
+                        _ilGen.Emit(OpCodes.Shr_Un);
+                        _ilGen.Emit(OpCodes.Ldc_I4, 0x3f);
+                        _ilGen.Emit(OpCodes.And);
+                        EmitSend();
+                        break;
+                }
+            }
+            else
+            {
+                // Branch.
+                bool onNotZero = ((opCode >> 4) & 1) != 0;
+
+                EmitLoadGprA(opCode);
+
+                Label lblSkip = _ilGen.DefineLabel();
+
+                if (onNotZero)
+                {
+                    _ilGen.Emit(OpCodes.Brfalse, lblSkip);
+                }
+                else
+                {
+                    _ilGen.Emit(OpCodes.Brtrue, lblSkip);
+                }
+
+                bool noDelays = (opCode & 0x20) != 0;
+
+                if (!noDelays)
+                {
+                    Emit(code, offset + 1, labels);
+                }
+
+                int target = offset + (opCode >> 14);
+
+                _ilGen.Emit(OpCodes.Br, labels[target]);
+
+                _ilGen.MarkLabel(lblSkip);
+            }
+        }
+
+        /// <summary>
+        /// Emits IL for a Arithmetic and Logic Unit instruction.
+        /// </summary>
+        /// <param name="opCode">Instruction to be translated</param>
+        /// <exception cref="InvalidOperationException">Throw when the instruction encoding is invalid</exception>
+        private void EmitAluOp(int opCode)
+        {
+            AluOperation op = (AluOperation)(opCode & 7);
+
+            switch (op)
+            {
+                case AluOperation.AluReg:
+                    EmitAluOp((AluRegOperation)((opCode >> 17) & 0x1f), opCode);
+                    break;
+
+                case AluOperation.AddImmediate:
+                    EmitLoadGprA(opCode);
+                    EmitLoadImm(opCode);
+                    _ilGen.Emit(OpCodes.Add);
+                    break;
+
+                case AluOperation.BitfieldReplace:
+                case AluOperation.BitfieldExtractLslImm:
+                case AluOperation.BitfieldExtractLslReg:
+                    int bfSrcBit = (opCode >> 17) & 0x1f;
+                    int bfSize = (opCode >> 22) & 0x1f;
+                    int bfDstBit = (opCode >> 27) & 0x1f;
+
+                    int bfMask = (1 << bfSize) - 1;
+
+                    switch (op)
+                    {
+                        case AluOperation.BitfieldReplace:
+                            EmitLoadGprB(opCode);
+                            _ilGen.Emit(OpCodes.Ldc_I4, bfSrcBit);
+                            _ilGen.Emit(OpCodes.Shr_Un);
+                            _ilGen.Emit(OpCodes.Ldc_I4, bfMask);
+                            _ilGen.Emit(OpCodes.And);
+                            _ilGen.Emit(OpCodes.Ldc_I4, bfDstBit);
+                            _ilGen.Emit(OpCodes.Shl);
+                            EmitLoadGprA(opCode);
+                            _ilGen.Emit(OpCodes.Ldc_I4, ~(bfMask << bfDstBit));
+                            _ilGen.Emit(OpCodes.And);
+                            _ilGen.Emit(OpCodes.Or);
+                            break;
+
+                        case AluOperation.BitfieldExtractLslImm:
+                            EmitLoadGprB(opCode);
+                            EmitLoadGprA(opCode);
+                            _ilGen.Emit(OpCodes.Shr_Un);
+                            _ilGen.Emit(OpCodes.Ldc_I4, bfMask);
+                            _ilGen.Emit(OpCodes.And);
+                            _ilGen.Emit(OpCodes.Ldc_I4, bfDstBit);
+                            _ilGen.Emit(OpCodes.Shl);
+                            break;
+
+                        case AluOperation.BitfieldExtractLslReg:
+                            EmitLoadGprB(opCode);
+                            _ilGen.Emit(OpCodes.Ldc_I4, bfSrcBit);
+                            _ilGen.Emit(OpCodes.Shr_Un);
+                            _ilGen.Emit(OpCodes.Ldc_I4, bfMask);
+                            _ilGen.Emit(OpCodes.And);
+                            EmitLoadGprA(opCode);
+                            _ilGen.Emit(OpCodes.Shl);
+                            break;
+                    }
+                    break;
+
+                case AluOperation.ReadImmediate:
+                    _ilGen.Emit(OpCodes.Ldarg_1);
+                    EmitLoadGprA(opCode);
+                    EmitLoadImm(opCode);
+                    _ilGen.Emit(OpCodes.Add);
+                    _ilGen.Emit(OpCodes.Call, typeof(MacroJitContext).GetMethod(nameof(MacroJitContext.Read)));
+                    break;
+
+                default:
+                    throw new InvalidOperationException($"Invalid operation \"{op}\" on instruction 0x{opCode:X8}.");
+            }
+        }
+
+        /// <summary>
+        /// Emits IL for a binary Arithmetic and Logic Unit instruction.
+        /// </summary>
+        /// <param name="aluOp">Arithmetic and Logic Unit instruction</param>
+        /// <param name="opCode">Raw instruction</param>
+        /// <exception cref="InvalidOperationException">Throw when the instruction encoding is invalid</exception>
+        private void EmitAluOp(AluRegOperation aluOp, int opCode)
+        {
+            switch (aluOp)
+            {
+                case AluRegOperation.Add:
+                    EmitLoadGprA(opCode);
+                    _ilGen.Emit(OpCodes.Conv_U8);
+                    EmitLoadGprB(opCode);
+                    _ilGen.Emit(OpCodes.Conv_U8);
+                    _ilGen.Emit(OpCodes.Add);
+                    _ilGen.Emit(OpCodes.Dup);
+                    _ilGen.Emit(OpCodes.Ldc_I8, 0xffffffffL);
+                    _ilGen.Emit(OpCodes.Cgt_Un);
+                    _ilGen.Emit(OpCodes.Stloc, _carry);
+                    _ilGen.Emit(OpCodes.Conv_U4);
+                    break;
+                case AluRegOperation.AddWithCarry:
+                    EmitLoadGprA(opCode);
+                    _ilGen.Emit(OpCodes.Conv_U8);
+                    EmitLoadGprB(opCode);
+                    _ilGen.Emit(OpCodes.Conv_U8);
+                    _ilGen.Emit(OpCodes.Ldloc_S, _carry);
+                    _ilGen.Emit(OpCodes.Conv_U8);
+                    _ilGen.Emit(OpCodes.Add);
+                    _ilGen.Emit(OpCodes.Add);
+                    _ilGen.Emit(OpCodes.Dup);
+                    _ilGen.Emit(OpCodes.Ldc_I8, 0xffffffffL);
+                    _ilGen.Emit(OpCodes.Cgt_Un);
+                    _ilGen.Emit(OpCodes.Stloc, _carry);
+                    _ilGen.Emit(OpCodes.Conv_U4);
+                    break;
+                case AluRegOperation.Subtract:
+                    EmitLoadGprA(opCode);
+                    _ilGen.Emit(OpCodes.Conv_U8);
+                    EmitLoadGprB(opCode);
+                    _ilGen.Emit(OpCodes.Conv_U8);
+                    _ilGen.Emit(OpCodes.Sub);
+                    _ilGen.Emit(OpCodes.Dup);
+                    _ilGen.Emit(OpCodes.Ldc_I8, 0x100000000L);
+                    _ilGen.Emit(OpCodes.Clt_Un);
+                    _ilGen.Emit(OpCodes.Stloc, _carry);
+                    _ilGen.Emit(OpCodes.Conv_U4);
+                    break;
+                case AluRegOperation.SubtractWithBorrow:
+                    EmitLoadGprA(opCode);
+                    _ilGen.Emit(OpCodes.Conv_U8);
+                    EmitLoadGprB(opCode);
+                    _ilGen.Emit(OpCodes.Conv_U8);
+                    _ilGen.Emit(OpCodes.Ldc_I4_1);
+                    _ilGen.Emit(OpCodes.Ldloc_S, _carry);
+                    _ilGen.Emit(OpCodes.Sub);
+                    _ilGen.Emit(OpCodes.Conv_U8);
+                    _ilGen.Emit(OpCodes.Sub);
+                    _ilGen.Emit(OpCodes.Sub);
+                    _ilGen.Emit(OpCodes.Dup);
+                    _ilGen.Emit(OpCodes.Ldc_I8, 0x100000000L);
+                    _ilGen.Emit(OpCodes.Clt_Un);
+                    _ilGen.Emit(OpCodes.Stloc, _carry);
+                    _ilGen.Emit(OpCodes.Conv_U4);
+                    break;
+                case AluRegOperation.BitwiseExclusiveOr:
+                    EmitLoadGprA(opCode);
+                    EmitLoadGprB(opCode);
+                    _ilGen.Emit(OpCodes.Xor);
+                    break;
+                case AluRegOperation.BitwiseOr:
+                    EmitLoadGprA(opCode);
+                    EmitLoadGprB(opCode);
+                    _ilGen.Emit(OpCodes.Or);
+                    break;
+                case AluRegOperation.BitwiseAnd:
+                    EmitLoadGprA(opCode);
+                    EmitLoadGprB(opCode);
+                    _ilGen.Emit(OpCodes.And);
+                    break;
+                case AluRegOperation.BitwiseAndNot:
+                    EmitLoadGprA(opCode);
+                    EmitLoadGprB(opCode);
+                    _ilGen.Emit(OpCodes.Not);
+                    _ilGen.Emit(OpCodes.And);
+                    break;
+                case AluRegOperation.BitwiseNotAnd:
+                    EmitLoadGprA(opCode);
+                    EmitLoadGprB(opCode);
+                    _ilGen.Emit(OpCodes.And);
+                    _ilGen.Emit(OpCodes.Not);
+                    break;
+                default:
+                    throw new InvalidOperationException($"Invalid operation \"{aluOp}\" on instruction 0x{opCode:X8}.");
+            }
+        }
+
+        /// <summary>
+        /// Loads a immediate value on the IL evaluation stack.
+        /// </summary>
+        /// <param name="opCode">Instruction from where the immediate should be extracted</param>
+        private void EmitLoadImm(int opCode)
+        {
+            // Note: The immediate is signed, the sign-extension is intended here.
+            _ilGen.Emit(OpCodes.Ldc_I4, opCode >> 14);
+        }
+
+        /// <summary>
+        /// Loads a value from the General Purpose register specified as first operand on the IL evaluation stack.
+        /// </summary>
+        /// <param name="opCode">Instruction from where the register number should be extracted</param>
+        private void EmitLoadGprA(int opCode)
+        {
+            EmitLoadGpr((opCode >> 11) & 7);
+        }
+
+        /// <summary>
+        /// Loads a value from the General Purpose register specified as second operand on the IL evaluation stack.
+        /// </summary>
+        /// <param name="opCode">Instruction from where the register number should be extracted</param>
+        private void EmitLoadGprB(int opCode)
+        {
+            EmitLoadGpr((opCode >> 14) & 7);
+        }
+
+        /// <summary>
+        /// Loads a value a General Purpose register on the IL evaluation stack.
+        /// </summary>
+        /// <remarks>
+        /// Register number 0 has a hardcoded value of 0.
+        /// </remarks>
+        /// <param name="index">Register number</param>
+        private void EmitLoadGpr(int index)
+        {
+            if (index == 0)
+            {
+                _ilGen.Emit(OpCodes.Ldc_I4_0);
+            }
+            else
+            {
+                _ilGen.Emit(OpCodes.Ldloc_S, _gprs[index]);
+            }
+        }
+
+        /// <summary>
+        /// Emits a call to the method that fetches an argument from the arguments FIFO.
+        /// The argument is pushed into the IL evaluation stack.
+        /// </summary>
+        private void EmitFetchParam()
+        {
+            _ilGen.Emit(OpCodes.Ldarg_0);
+            _ilGen.Emit(OpCodes.Call, typeof(MacroJitContext).GetMethod(nameof(MacroJitContext.FetchParam)));
+        }
+
+        /// <summary>
+        /// Stores the value on the top of the IL evaluation stack into a General Purpose register.
+        /// </summary>
+        /// <remarks>
+        /// Register number 0 does not exist, reads are hardcoded to 0, and writes are simply discarded.
+        /// </remarks>
+        /// <param name="opCode">Instruction from where the register number should be extracted</param>
+        private void EmitStoreDstGpr(int opCode)
+        {
+            int index = (opCode >> 8) & 7;
+
+            if (index == 0)
+            {
+                _ilGen.Emit(OpCodes.Pop);
+            }
+            else
+            {
+                _ilGen.Emit(OpCodes.Stloc_S, _gprs[index]);
+            }
+        }
+
+        /// <summary>
+        /// Stores the value on the top of the IL evaluation stack as method address.
+        /// This will be used on subsequent send calls as the destination method address.
+        /// Additionally, the 6 bits starting at bit 12 will be used as increment value,
+        /// added to the method address after each sent value.
+        /// </summary>
+        private void EmitStoreMethAddr()
+        {
+            _ilGen.Emit(OpCodes.Dup);
+            _ilGen.Emit(OpCodes.Ldc_I4, 0xfff);
+            _ilGen.Emit(OpCodes.And);
+            _ilGen.Emit(OpCodes.Stloc_S, _methAddr);
+            _ilGen.Emit(OpCodes.Ldc_I4, 12);
+            _ilGen.Emit(OpCodes.Shr_Un);
+            _ilGen.Emit(OpCodes.Ldc_I4, 0x3f);
+            _ilGen.Emit(OpCodes.And);
+            _ilGen.Emit(OpCodes.Stloc_S, _methIncr);
+        }
+
+        /// <summary>
+        /// Sends the value on the top of the IL evaluation stack to the GPU,
+        /// using the current method address.
+        /// </summary>
+        private void EmitSend()
+        {
+            _ilGen.Emit(OpCodes.Ldarg_1);
+            _ilGen.Emit(OpCodes.Ldloc_S, _methAddr);
+            _ilGen.Emit(OpCodes.Call, typeof(MacroJitContext).GetMethod(nameof(MacroJitContext.Send)));
+            _ilGen.Emit(OpCodes.Ldloc_S, _methAddr);
+            _ilGen.Emit(OpCodes.Ldloc_S, _methIncr);
+            _ilGen.Emit(OpCodes.Add);
+            _ilGen.Emit(OpCodes.Stloc_S, _methAddr);
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroJitContext.cs b/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroJitContext.cs
new file mode 100644
index 00000000..52c2a11b
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/MME/MacroJitContext.cs
@@ -0,0 +1,55 @@
+using Ryujinx.Common.Logging;
+using Ryujinx.Graphics.Device;
+using System.Collections.Generic;
+
+namespace Ryujinx.Graphics.Gpu.Engine.MME
+{
+    /// <summary>
+    /// Represents a Macro Just-in-Time compiler execution context.
+    /// </summary>
+    class MacroJitContext
+    {
+        /// <summary>
+        /// Arguments FIFO.
+        /// </summary>
+        public Queue<FifoWord> Fifo { get; } = new Queue<FifoWord>();
+
+        /// <summary>
+        /// Fetches a arguments from the arguments FIFO.
+        /// </summary>
+        /// <returns>The call argument, or 0 if the FIFO is empty</returns>
+        public int FetchParam()
+        {
+            if (!Fifo.TryDequeue(out var value))
+            {
+                Logger.Warning?.Print(LogClass.Gpu, "Macro attempted to fetch an inexistent argument.");
+
+                return 0;
+            }
+
+            return value.Word;
+        }
+
+        /// <summary>
+        /// Reads data from a GPU register.
+        /// </summary>
+        /// <param name="state">Current GPU state</param>
+        /// <param name="reg">Register offset to read</param>
+        /// <returns>GPU register value</returns>
+        public static int Read(IDeviceState state, int reg)
+        {
+            return state.Read(reg * 4);
+        }
+
+        /// <summary>
+        /// Performs a GPU method call.
+        /// </summary>
+        /// <param name="value">Call argument</param>
+        /// <param name="state">Current GPU state</param>
+        /// <param name="methAddr">Address, in words, of the method</param>
+        public static void Send(int value, IDeviceState state, int methAddr)
+        {
+            state.Write(methAddr * 4, value);
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/MmeShadowScratch.cs b/src/Ryujinx.Graphics.Gpu/Engine/MmeShadowScratch.cs
new file mode 100644
index 00000000..44cd8213
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/MmeShadowScratch.cs
@@ -0,0 +1,18 @@
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    /// <summary>
+    /// Represents temporary storage used by macros.
+    /// </summary>
+    [StructLayout(LayoutKind.Sequential, Size = 1024)]
+    struct MmeShadowScratch
+    {
+#pragma warning disable CS0169
+        private uint _e0;
+#pragma warning restore CS0169
+        public ref uint this[int index] => ref AsSpan()[index];
+        public Span<uint> AsSpan() => MemoryMarshal.CreateSpan(ref _e0, 256);
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/SetMmeShadowRamControlMode.cs b/src/Ryujinx.Graphics.Gpu/Engine/SetMmeShadowRamControlMode.cs
new file mode 100644
index 00000000..060d35ca
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/SetMmeShadowRamControlMode.cs
@@ -0,0 +1,13 @@
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    /// <summary>
+    /// MME shadow RAM control mode.
+    /// </summary>
+    enum SetMmeShadowRamControlMode
+    {
+        MethodTrack = 0,
+        MethodTrackWithFilter = 1,
+        MethodPassthrough = 2,
+        MethodReplay = 3,
+    }
+}
\ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/ShaderTexture.cs b/src/Ryujinx.Graphics.Gpu/Engine/ShaderTexture.cs
new file mode 100644
index 00000000..e1e3085b
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/ShaderTexture.cs
@@ -0,0 +1,111 @@
+using Ryujinx.Common.Logging;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Shader;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    /// <summary>
+    /// Shader texture properties conversion methods.
+    /// </summary>
+    static class ShaderTexture
+    {
+        /// <summary>
+        /// Gets a texture target from a sampler type.
+        /// </summary>
+        /// <param name="type">Sampler type</param>
+        /// <returns>Texture target value</returns>
+        public static Target GetTarget(SamplerType type)
+        {
+            type &= ~(SamplerType.Indexed | SamplerType.Shadow);
+
+            switch (type)
+            {
+                case SamplerType.Texture1D:
+                    return Target.Texture1D;
+
+                case SamplerType.TextureBuffer:
+                    return Target.TextureBuffer;
+
+                case SamplerType.Texture1D | SamplerType.Array:
+                    return Target.Texture1DArray;
+
+                case SamplerType.Texture2D:
+                    return Target.Texture2D;
+
+                case SamplerType.Texture2D | SamplerType.Array:
+                    return Target.Texture2DArray;
+
+                case SamplerType.Texture2D | SamplerType.Multisample:
+                    return Target.Texture2DMultisample;
+
+                case SamplerType.Texture2D | SamplerType.Multisample | SamplerType.Array:
+                    return Target.Texture2DMultisampleArray;
+
+                case SamplerType.Texture3D:
+                    return Target.Texture3D;
+
+                case SamplerType.TextureCube:
+                    return Target.Cubemap;
+
+                case SamplerType.TextureCube | SamplerType.Array:
+                    return Target.CubemapArray;
+            }
+
+            Logger.Warning?.Print(LogClass.Gpu, $"Invalid sampler type \"{type}\".");
+
+            return Target.Texture2D;
+        }
+
+        /// <summary>
+        /// Gets a texture format from a shader image format.
+        /// </summary>
+        /// <param name="format">Shader image format</param>
+        /// <returns>Texture format</returns>
+        public static Format GetFormat(TextureFormat format)
+        {
+            return format switch
+            {
+                TextureFormat.R8Unorm           => Format.R8Unorm,
+                TextureFormat.R8Snorm           => Format.R8Snorm,
+                TextureFormat.R8Uint            => Format.R8Uint,
+                TextureFormat.R8Sint            => Format.R8Sint,
+                TextureFormat.R16Float          => Format.R16Float,
+                TextureFormat.R16Unorm          => Format.R16Unorm,
+                TextureFormat.R16Snorm          => Format.R16Snorm,
+                TextureFormat.R16Uint           => Format.R16Uint,
+                TextureFormat.R16Sint           => Format.R16Sint,
+                TextureFormat.R32Float          => Format.R32Float,
+                TextureFormat.R32Uint           => Format.R32Uint,
+                TextureFormat.R32Sint           => Format.R32Sint,
+                TextureFormat.R8G8Unorm         => Format.R8G8Unorm,
+                TextureFormat.R8G8Snorm         => Format.R8G8Snorm,
+                TextureFormat.R8G8Uint          => Format.R8G8Uint,
+                TextureFormat.R8G8Sint          => Format.R8G8Sint,
+                TextureFormat.R16G16Float       => Format.R16G16Float,
+                TextureFormat.R16G16Unorm       => Format.R16G16Unorm,
+                TextureFormat.R16G16Snorm       => Format.R16G16Snorm,
+                TextureFormat.R16G16Uint        => Format.R16G16Uint,
+                TextureFormat.R16G16Sint        => Format.R16G16Sint,
+                TextureFormat.R32G32Float       => Format.R32G32Float,
+                TextureFormat.R32G32Uint        => Format.R32G32Uint,
+                TextureFormat.R32G32Sint        => Format.R32G32Sint,
+                TextureFormat.R8G8B8A8Unorm     => Format.R8G8B8A8Unorm,
+                TextureFormat.R8G8B8A8Snorm     => Format.R8G8B8A8Snorm,
+                TextureFormat.R8G8B8A8Uint      => Format.R8G8B8A8Uint,
+                TextureFormat.R8G8B8A8Sint      => Format.R8G8B8A8Sint,
+                TextureFormat.R16G16B16A16Float => Format.R16G16B16A16Float,
+                TextureFormat.R16G16B16A16Unorm => Format.R16G16B16A16Unorm,
+                TextureFormat.R16G16B16A16Snorm => Format.R16G16B16A16Snorm,
+                TextureFormat.R16G16B16A16Uint  => Format.R16G16B16A16Uint,
+                TextureFormat.R16G16B16A16Sint  => Format.R16G16B16A16Sint,
+                TextureFormat.R32G32B32A32Float => Format.R32G32B32A32Float,
+                TextureFormat.R32G32B32A32Uint  => Format.R32G32B32A32Uint,
+                TextureFormat.R32G32B32A32Sint  => Format.R32G32B32A32Sint,
+                TextureFormat.R10G10B10A2Unorm  => Format.R10G10B10A2Unorm,
+                TextureFormat.R10G10B10A2Uint   => Format.R10G10B10A2Uint,
+                TextureFormat.R11G11B10Float    => Format.R11G11B10Float,
+                _                               => 0
+            };
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendFunctions.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendFunctions.cs
new file mode 100644
index 00000000..a40b9cc4
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendFunctions.cs
@@ -0,0 +1,4226 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.GAL;
+using System.Globalization;
+using System.Runtime.InteropServices;
+using System.Text;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed.Blender
+{
+    static class AdvancedBlendFunctions
+    {
+        public static readonly AdvancedBlendUcode[] Table = new AdvancedBlendUcode[]
+        {
+            new AdvancedBlendUcode(AdvancedBlendOp.PlusClamped,      AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedPlusClampedPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.PlusClampedAlpha, AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedPlusClampedAlphaPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.PlusDarker,       AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedPlusDarkerPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Multiply,         AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedMultiplyPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Screen,           AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedScreenPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Overlay,          AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedOverlayPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Darken,           AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedDarkenPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Lighten,          AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedLightenPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.ColorDodge,       AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedColorDodgePremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.ColorBurn,        AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedColorBurnPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HardLight,        AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedHardLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.SoftLight,        AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedSoftLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Difference,       AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedDifferencePremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Minus,            AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedMinusPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.MinusClamped,     AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedMinusClampedPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Exclusion,        AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedExclusionPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Contrast,         AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedContrastPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Invert,           AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedInvertPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.InvertRGB,        AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedInvertRGBPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.InvertOvg,        AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedInvertOvgPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearDodge,      AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedLinearDodgePremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearBurn,       AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedLinearBurnPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.VividLight,       AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedVividLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearLight,      AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedLinearLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.PinLight,         AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedPinLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HardMix,          AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedHardMixPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Red,              AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedRedPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Green,            AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedGreenPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Blue,             AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedBluePremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslHue,           AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedHslHuePremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslSaturation,    AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedHslSaturationPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslColor,         AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedHslColorPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslLuminosity,    AdvancedBlendOverlap.Uncorrelated, true,  GenUncorrelatedHslLuminosityPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Src,              AdvancedBlendOverlap.Disjoint,     true,  GenDisjointSrcPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Dst,              AdvancedBlendOverlap.Disjoint,     true,  GenDisjointDstPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcOver,          AdvancedBlendOverlap.Disjoint,     true,  GenDisjointSrcOverPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.DstOver,          AdvancedBlendOverlap.Disjoint,     true,  GenDisjointDstOverPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcIn,            AdvancedBlendOverlap.Disjoint,     true,  GenDisjointSrcInPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.DstIn,            AdvancedBlendOverlap.Disjoint,     true,  GenDisjointDstInPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcOut,           AdvancedBlendOverlap.Disjoint,     true,  GenDisjointSrcOutPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.DstOut,           AdvancedBlendOverlap.Disjoint,     true,  GenDisjointDstOutPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcAtop,          AdvancedBlendOverlap.Disjoint,     true,  GenDisjointSrcAtopPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.DstAtop,          AdvancedBlendOverlap.Disjoint,     true,  GenDisjointDstAtopPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Xor,              AdvancedBlendOverlap.Disjoint,     true,  GenDisjointXorPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Plus,             AdvancedBlendOverlap.Disjoint,     true,  GenDisjointPlusPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Multiply,         AdvancedBlendOverlap.Disjoint,     true,  GenDisjointMultiplyPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Screen,           AdvancedBlendOverlap.Disjoint,     true,  GenDisjointScreenPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Overlay,          AdvancedBlendOverlap.Disjoint,     true,  GenDisjointOverlayPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Darken,           AdvancedBlendOverlap.Disjoint,     true,  GenDisjointDarkenPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Lighten,          AdvancedBlendOverlap.Disjoint,     true,  GenDisjointLightenPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.ColorDodge,       AdvancedBlendOverlap.Disjoint,     true,  GenDisjointColorDodgePremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.ColorBurn,        AdvancedBlendOverlap.Disjoint,     true,  GenDisjointColorBurnPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HardLight,        AdvancedBlendOverlap.Disjoint,     true,  GenDisjointHardLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.SoftLight,        AdvancedBlendOverlap.Disjoint,     true,  GenDisjointSoftLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Difference,       AdvancedBlendOverlap.Disjoint,     true,  GenDisjointDifferencePremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Exclusion,        AdvancedBlendOverlap.Disjoint,     true,  GenDisjointExclusionPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Invert,           AdvancedBlendOverlap.Disjoint,     true,  GenDisjointInvertPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.InvertRGB,        AdvancedBlendOverlap.Disjoint,     true,  GenDisjointInvertRGBPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearDodge,      AdvancedBlendOverlap.Disjoint,     true,  GenDisjointLinearDodgePremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearBurn,       AdvancedBlendOverlap.Disjoint,     true,  GenDisjointLinearBurnPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.VividLight,       AdvancedBlendOverlap.Disjoint,     true,  GenDisjointVividLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearLight,      AdvancedBlendOverlap.Disjoint,     true,  GenDisjointLinearLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.PinLight,         AdvancedBlendOverlap.Disjoint,     true,  GenDisjointPinLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HardMix,          AdvancedBlendOverlap.Disjoint,     true,  GenDisjointHardMixPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslHue,           AdvancedBlendOverlap.Disjoint,     true,  GenDisjointHslHuePremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslSaturation,    AdvancedBlendOverlap.Disjoint,     true,  GenDisjointHslSaturationPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslColor,         AdvancedBlendOverlap.Disjoint,     true,  GenDisjointHslColorPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslLuminosity,    AdvancedBlendOverlap.Disjoint,     true,  GenDisjointHslLuminosityPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Src,              AdvancedBlendOverlap.Conjoint,     true,  GenConjointSrcPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Dst,              AdvancedBlendOverlap.Conjoint,     true,  GenConjointDstPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcOver,          AdvancedBlendOverlap.Conjoint,     true,  GenConjointSrcOverPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.DstOver,          AdvancedBlendOverlap.Conjoint,     true,  GenConjointDstOverPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcIn,            AdvancedBlendOverlap.Conjoint,     true,  GenConjointSrcInPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.DstIn,            AdvancedBlendOverlap.Conjoint,     true,  GenConjointDstInPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcOut,           AdvancedBlendOverlap.Conjoint,     true,  GenConjointSrcOutPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.DstOut,           AdvancedBlendOverlap.Conjoint,     true,  GenConjointDstOutPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcAtop,          AdvancedBlendOverlap.Conjoint,     true,  GenConjointSrcAtopPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.DstAtop,          AdvancedBlendOverlap.Conjoint,     true,  GenConjointDstAtopPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Xor,              AdvancedBlendOverlap.Conjoint,     true,  GenConjointXorPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Multiply,         AdvancedBlendOverlap.Conjoint,     true,  GenConjointMultiplyPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Screen,           AdvancedBlendOverlap.Conjoint,     true,  GenConjointScreenPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Overlay,          AdvancedBlendOverlap.Conjoint,     true,  GenConjointOverlayPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Darken,           AdvancedBlendOverlap.Conjoint,     true,  GenConjointDarkenPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Lighten,          AdvancedBlendOverlap.Conjoint,     true,  GenConjointLightenPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.ColorDodge,       AdvancedBlendOverlap.Conjoint,     true,  GenConjointColorDodgePremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.ColorBurn,        AdvancedBlendOverlap.Conjoint,     true,  GenConjointColorBurnPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HardLight,        AdvancedBlendOverlap.Conjoint,     true,  GenConjointHardLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.SoftLight,        AdvancedBlendOverlap.Conjoint,     true,  GenConjointSoftLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Difference,       AdvancedBlendOverlap.Conjoint,     true,  GenConjointDifferencePremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Exclusion,        AdvancedBlendOverlap.Conjoint,     true,  GenConjointExclusionPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.Invert,           AdvancedBlendOverlap.Conjoint,     true,  GenConjointInvertPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.InvertRGB,        AdvancedBlendOverlap.Conjoint,     true,  GenConjointInvertRGBPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearDodge,      AdvancedBlendOverlap.Conjoint,     true,  GenConjointLinearDodgePremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearBurn,       AdvancedBlendOverlap.Conjoint,     true,  GenConjointLinearBurnPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.VividLight,       AdvancedBlendOverlap.Conjoint,     true,  GenConjointVividLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearLight,      AdvancedBlendOverlap.Conjoint,     true,  GenConjointLinearLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.PinLight,         AdvancedBlendOverlap.Conjoint,     true,  GenConjointPinLightPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HardMix,          AdvancedBlendOverlap.Conjoint,     true,  GenConjointHardMixPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslHue,           AdvancedBlendOverlap.Conjoint,     true,  GenConjointHslHuePremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslSaturation,    AdvancedBlendOverlap.Conjoint,     true,  GenConjointHslSaturationPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslColor,         AdvancedBlendOverlap.Conjoint,     true,  GenConjointHslColorPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslLuminosity,    AdvancedBlendOverlap.Conjoint,     true,  GenConjointHslLuminosityPremul),
+            new AdvancedBlendUcode(AdvancedBlendOp.DstOver,          AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedDstOver),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcIn,            AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedSrcIn),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcOut,           AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedSrcOut),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcAtop,          AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedSrcAtop),
+            new AdvancedBlendUcode(AdvancedBlendOp.DstAtop,          AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedDstAtop),
+            new AdvancedBlendUcode(AdvancedBlendOp.Xor,              AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedXor),
+            new AdvancedBlendUcode(AdvancedBlendOp.PlusClamped,      AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedPlusClamped),
+            new AdvancedBlendUcode(AdvancedBlendOp.PlusClampedAlpha, AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedPlusClampedAlpha),
+            new AdvancedBlendUcode(AdvancedBlendOp.PlusDarker,       AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedPlusDarker),
+            new AdvancedBlendUcode(AdvancedBlendOp.Multiply,         AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedMultiply),
+            new AdvancedBlendUcode(AdvancedBlendOp.Screen,           AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedScreen),
+            new AdvancedBlendUcode(AdvancedBlendOp.Overlay,          AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedOverlay),
+            new AdvancedBlendUcode(AdvancedBlendOp.Darken,           AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedDarken),
+            new AdvancedBlendUcode(AdvancedBlendOp.Lighten,          AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedLighten),
+            new AdvancedBlendUcode(AdvancedBlendOp.ColorDodge,       AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedColorDodge),
+            new AdvancedBlendUcode(AdvancedBlendOp.ColorBurn,        AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedColorBurn),
+            new AdvancedBlendUcode(AdvancedBlendOp.HardLight,        AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedHardLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.SoftLight,        AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedSoftLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.Difference,       AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedDifference),
+            new AdvancedBlendUcode(AdvancedBlendOp.Minus,            AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedMinus),
+            new AdvancedBlendUcode(AdvancedBlendOp.MinusClamped,     AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedMinusClamped),
+            new AdvancedBlendUcode(AdvancedBlendOp.Exclusion,        AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedExclusion),
+            new AdvancedBlendUcode(AdvancedBlendOp.Contrast,         AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedContrast),
+            new AdvancedBlendUcode(AdvancedBlendOp.InvertRGB,        AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedInvertRGB),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearDodge,      AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedLinearDodge),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearBurn,       AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedLinearBurn),
+            new AdvancedBlendUcode(AdvancedBlendOp.VividLight,       AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedVividLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearLight,      AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedLinearLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.PinLight,         AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedPinLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.HardMix,          AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedHardMix),
+            new AdvancedBlendUcode(AdvancedBlendOp.Red,              AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedRed),
+            new AdvancedBlendUcode(AdvancedBlendOp.Green,            AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedGreen),
+            new AdvancedBlendUcode(AdvancedBlendOp.Blue,             AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedBlue),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslHue,           AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedHslHue),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslSaturation,    AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedHslSaturation),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslColor,         AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedHslColor),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslLuminosity,    AdvancedBlendOverlap.Uncorrelated, false, GenUncorrelatedHslLuminosity),
+            new AdvancedBlendUcode(AdvancedBlendOp.Src,              AdvancedBlendOverlap.Disjoint,     false, GenDisjointSrc),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcOver,          AdvancedBlendOverlap.Disjoint,     false, GenDisjointSrcOver),
+            new AdvancedBlendUcode(AdvancedBlendOp.DstOver,          AdvancedBlendOverlap.Disjoint,     false, GenDisjointDstOver),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcIn,            AdvancedBlendOverlap.Disjoint,     false, GenDisjointSrcIn),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcOut,           AdvancedBlendOverlap.Disjoint,     false, GenDisjointSrcOut),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcAtop,          AdvancedBlendOverlap.Disjoint,     false, GenDisjointSrcAtop),
+            new AdvancedBlendUcode(AdvancedBlendOp.DstAtop,          AdvancedBlendOverlap.Disjoint,     false, GenDisjointDstAtop),
+            new AdvancedBlendUcode(AdvancedBlendOp.Xor,              AdvancedBlendOverlap.Disjoint,     false, GenDisjointXor),
+            new AdvancedBlendUcode(AdvancedBlendOp.Plus,             AdvancedBlendOverlap.Disjoint,     false, GenDisjointPlus),
+            new AdvancedBlendUcode(AdvancedBlendOp.Multiply,         AdvancedBlendOverlap.Disjoint,     false, GenDisjointMultiply),
+            new AdvancedBlendUcode(AdvancedBlendOp.Screen,           AdvancedBlendOverlap.Disjoint,     false, GenDisjointScreen),
+            new AdvancedBlendUcode(AdvancedBlendOp.Overlay,          AdvancedBlendOverlap.Disjoint,     false, GenDisjointOverlay),
+            new AdvancedBlendUcode(AdvancedBlendOp.Darken,           AdvancedBlendOverlap.Disjoint,     false, GenDisjointDarken),
+            new AdvancedBlendUcode(AdvancedBlendOp.Lighten,          AdvancedBlendOverlap.Disjoint,     false, GenDisjointLighten),
+            new AdvancedBlendUcode(AdvancedBlendOp.ColorDodge,       AdvancedBlendOverlap.Disjoint,     false, GenDisjointColorDodge),
+            new AdvancedBlendUcode(AdvancedBlendOp.ColorBurn,        AdvancedBlendOverlap.Disjoint,     false, GenDisjointColorBurn),
+            new AdvancedBlendUcode(AdvancedBlendOp.HardLight,        AdvancedBlendOverlap.Disjoint,     false, GenDisjointHardLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.SoftLight,        AdvancedBlendOverlap.Disjoint,     false, GenDisjointSoftLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.Difference,       AdvancedBlendOverlap.Disjoint,     false, GenDisjointDifference),
+            new AdvancedBlendUcode(AdvancedBlendOp.Exclusion,        AdvancedBlendOverlap.Disjoint,     false, GenDisjointExclusion),
+            new AdvancedBlendUcode(AdvancedBlendOp.InvertRGB,        AdvancedBlendOverlap.Disjoint,     false, GenDisjointInvertRGB),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearDodge,      AdvancedBlendOverlap.Disjoint,     false, GenDisjointLinearDodge),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearBurn,       AdvancedBlendOverlap.Disjoint,     false, GenDisjointLinearBurn),
+            new AdvancedBlendUcode(AdvancedBlendOp.VividLight,       AdvancedBlendOverlap.Disjoint,     false, GenDisjointVividLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearLight,      AdvancedBlendOverlap.Disjoint,     false, GenDisjointLinearLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.PinLight,         AdvancedBlendOverlap.Disjoint,     false, GenDisjointPinLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.HardMix,          AdvancedBlendOverlap.Disjoint,     false, GenDisjointHardMix),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslHue,           AdvancedBlendOverlap.Disjoint,     false, GenDisjointHslHue),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslSaturation,    AdvancedBlendOverlap.Disjoint,     false, GenDisjointHslSaturation),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslColor,         AdvancedBlendOverlap.Disjoint,     false, GenDisjointHslColor),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslLuminosity,    AdvancedBlendOverlap.Disjoint,     false, GenDisjointHslLuminosity),
+            new AdvancedBlendUcode(AdvancedBlendOp.Src,              AdvancedBlendOverlap.Conjoint,     false, GenConjointSrc),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcOver,          AdvancedBlendOverlap.Conjoint,     false, GenConjointSrcOver),
+            new AdvancedBlendUcode(AdvancedBlendOp.DstOver,          AdvancedBlendOverlap.Conjoint,     false, GenConjointDstOver),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcIn,            AdvancedBlendOverlap.Conjoint,     false, GenConjointSrcIn),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcOut,           AdvancedBlendOverlap.Conjoint,     false, GenConjointSrcOut),
+            new AdvancedBlendUcode(AdvancedBlendOp.SrcAtop,          AdvancedBlendOverlap.Conjoint,     false, GenConjointSrcAtop),
+            new AdvancedBlendUcode(AdvancedBlendOp.DstAtop,          AdvancedBlendOverlap.Conjoint,     false, GenConjointDstAtop),
+            new AdvancedBlendUcode(AdvancedBlendOp.Xor,              AdvancedBlendOverlap.Conjoint,     false, GenConjointXor),
+            new AdvancedBlendUcode(AdvancedBlendOp.Multiply,         AdvancedBlendOverlap.Conjoint,     false, GenConjointMultiply),
+            new AdvancedBlendUcode(AdvancedBlendOp.Screen,           AdvancedBlendOverlap.Conjoint,     false, GenConjointScreen),
+            new AdvancedBlendUcode(AdvancedBlendOp.Overlay,          AdvancedBlendOverlap.Conjoint,     false, GenConjointOverlay),
+            new AdvancedBlendUcode(AdvancedBlendOp.Darken,           AdvancedBlendOverlap.Conjoint,     false, GenConjointDarken),
+            new AdvancedBlendUcode(AdvancedBlendOp.Lighten,          AdvancedBlendOverlap.Conjoint,     false, GenConjointLighten),
+            new AdvancedBlendUcode(AdvancedBlendOp.ColorDodge,       AdvancedBlendOverlap.Conjoint,     false, GenConjointColorDodge),
+            new AdvancedBlendUcode(AdvancedBlendOp.ColorBurn,        AdvancedBlendOverlap.Conjoint,     false, GenConjointColorBurn),
+            new AdvancedBlendUcode(AdvancedBlendOp.HardLight,        AdvancedBlendOverlap.Conjoint,     false, GenConjointHardLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.SoftLight,        AdvancedBlendOverlap.Conjoint,     false, GenConjointSoftLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.Difference,       AdvancedBlendOverlap.Conjoint,     false, GenConjointDifference),
+            new AdvancedBlendUcode(AdvancedBlendOp.Exclusion,        AdvancedBlendOverlap.Conjoint,     false, GenConjointExclusion),
+            new AdvancedBlendUcode(AdvancedBlendOp.InvertRGB,        AdvancedBlendOverlap.Conjoint,     false, GenConjointInvertRGB),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearDodge,      AdvancedBlendOverlap.Conjoint,     false, GenConjointLinearDodge),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearBurn,       AdvancedBlendOverlap.Conjoint,     false, GenConjointLinearBurn),
+            new AdvancedBlendUcode(AdvancedBlendOp.VividLight,       AdvancedBlendOverlap.Conjoint,     false, GenConjointVividLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.LinearLight,      AdvancedBlendOverlap.Conjoint,     false, GenConjointLinearLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.PinLight,         AdvancedBlendOverlap.Conjoint,     false, GenConjointPinLight),
+            new AdvancedBlendUcode(AdvancedBlendOp.HardMix,          AdvancedBlendOverlap.Conjoint,     false, GenConjointHardMix),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslHue,           AdvancedBlendOverlap.Conjoint,     false, GenConjointHslHue),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslSaturation,    AdvancedBlendOverlap.Conjoint,     false, GenConjointHslSaturation),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslColor,         AdvancedBlendOverlap.Conjoint,     false, GenConjointHslColor),
+            new AdvancedBlendUcode(AdvancedBlendOp.HslLuminosity,    AdvancedBlendOverlap.Conjoint,     false, GenConjointHslLuminosity)
+        };
+
+        public static string GenTable()
+        {
+            // This can be used to generate the table on AdvancedBlendPreGenTable.
+
+            StringBuilder sb = new StringBuilder();
+
+            sb.AppendLine($"private static Dictionary<Hash128, AdvancedBlendEntry> _entries = new()");
+            sb.AppendLine("{");
+
+            foreach (var entry in Table)
+            {
+                Hash128 hash = XXHash128.ComputeHash(MemoryMarshal.Cast<uint, byte>(entry.Code));
+
+                string[] constants = new string[entry.Constants != null ? entry.Constants.Length : 0];
+
+                for (int i = 0; i < constants.Length; i++)
+                {
+                    RgbFloat rgb = entry.Constants[i];
+
+                    constants[i] = string.Format(CultureInfo.InvariantCulture, "new " + nameof(RgbFloat) + "({0}f, {1}f, {2}f)", rgb.R, rgb.G, rgb.B);
+                }
+
+                string constantList = constants.Length > 0 ? $"new[] {{ {string.Join(", ", constants)} }}" : $"Array.Empty<{nameof(RgbFloat)}>()";
+
+                static string EnumValue(string name, object value)
+                {
+                    if (value.ToString() == "0")
+                    {
+                        return "0";
+                    }
+
+                    return $"{name}.{value}";
+                }
+
+                string alpha = $"new {nameof(FixedFunctionAlpha)}({EnumValue(nameof(BlendUcodeEnable), entry.Alpha.Enable)}, {EnumValue(nameof(BlendOp), entry.Alpha.AlphaOp)}, {EnumValue(nameof(BlendFactor), entry.Alpha.AlphaSrcFactor)}, {EnumValue(nameof(BlendFactor), entry.Alpha.AlphaDstFactor)})";
+
+                sb.AppendLine($"    {{ new Hash128(0x{hash.Low:X16}, 0x{hash.High:X16}), new AdvancedBlendEntry({nameof(AdvancedBlendOp)}.{entry.Op}, {nameof(AdvancedBlendOverlap)}.{entry.Overlap}, {(entry.SrcPreMultiplied ? "true" : "false")}, {constantList}, {alpha}) }},");
+            }
+
+            sb.AppendLine("};");
+
+            return sb.ToString();
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedPlusClampedPremul(ref UcodeAssembler asm)
+        {
+            asm.Add(CC.T, Dest.PBR, OpBD.DstRGB, OpBD.SrcRGB);
+            asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedPlusClampedAlphaPremul(ref UcodeAssembler asm)
+        {
+            asm.Add(CC.T, Dest.Temp0, OpBD.DstRGB, OpBD.SrcRGB);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+            asm.Min(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedPlusDarkerPremul(ref UcodeAssembler asm)
+        {
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+            asm.Add(CC.T, Dest.PBR, OpBD.PBR, OpBD.SrcRGB);
+            asm.Add(CC.T, Dest.PBR, OpBD.PBR, OpBD.DstRGB);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.SrcAAA);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.DstAAA);
+            asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedMultiplyPremul(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstRGB);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedScreenPremul(ref UcodeAssembler asm)
+        {
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.DstRGB);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedOverlayPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantRGB);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp2, OpBD.Temp1, OpAC.Temp2, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedDarkenPremul(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.SrcAAA);
+            asm.Min(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedLightenPremul(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.SrcAAA);
+            asm.Max(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedColorDodgePremul(ref UcodeAssembler asm)
+        {
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.SrcRGB);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.SrcAAA);
+            asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.DstRGB);
+            asm.Min(CC.GT, Dest.PBR, OpAC.DstAAA, OpBD.PBR);
+            asm.Mul(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.SrcAAA);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.DstRGB, OpBD.ConstantZero);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedColorBurnPremul(ref UcodeAssembler asm)
+        {
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.DstAAA, OpBD.SrcAAA, OpAC.SrcAAA, OpBD.DstRGB);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcRGB);
+            asm.Mul(CC.T, Dest.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.PBR);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.SrcAAA, OpBD.DstAAA, OpAC.SrcAAA, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantZero);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.DstAAA, OpBD.DstRGB);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedHardLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.Temp2, OpBD.ConstantRGB);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp2, OpBD.Temp1, OpAC.Temp2, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedSoftLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(4, 0.25f, 0.25f, 0.25f);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantRGB);
+            asm.SetConstant(0, 0.2605f, 0.2605f, 0.2605f);
+            asm.Mul(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(1, -0.7817f, -0.7817f, -0.7817f);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(2, 0.3022f, 0.3022f, 0.3022f);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(3, 0.2192f, 0.2192f, 0.2192f);
+            asm.Add(CC.GT, Dest.Temp0, OpBD.PBR, OpBD.ConstantRGB);
+            asm.SetConstant(5, 16f, 16f, 16f);
+            asm.Mul(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(6, 12f, 12f, 12f);
+            asm.Mmsub(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(7, 3f, 3f, 3f);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Mmsub(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.ConstantOne, OpAC.Temp1, OpBD.Temp1);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedDifferencePremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.Temp0, OpBD.Temp2, OpBD.Temp1);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedMinusPremul(ref UcodeAssembler asm)
+        {
+            asm.Sub(CC.T, Dest.Temp0, OpBD.DstRGB, OpBD.SrcRGB);
+            return new FixedFunctionAlpha(BlendOp.ReverseSubtractGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedMinusClampedPremul(ref UcodeAssembler asm)
+        {
+            asm.Sub(CC.T, Dest.PBR, OpBD.DstRGB, OpBD.SrcRGB);
+            asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Max(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantZero);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedExclusionPremul(ref UcodeAssembler asm)
+        {
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.DstRGB);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.DstRGB);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedContrastPremul(ref UcodeAssembler asm)
+        {
+            asm.SetConstant(0, 2f, 2f, 2f);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.ConstantRGB, OpAC.DstAAA, OpBD.ConstantOne);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.SrcAAA, OpBD.ConstantOne);
+            asm.Mul(CC.T, Dest.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.PBR, OpBD.DstAAA);
+            asm.SetConstant(1, 0.5f, 0.5f, 0.5f);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantRGB);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedInvertPremul(ref UcodeAssembler asm)
+        {
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA, OpAC.SrcAAA, OpBD.DstRGB);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.OneMinusSrcAAA, OpAC.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedInvertRGBPremul(ref UcodeAssembler asm)
+        {
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.DstAAA, OpAC.SrcRGB, OpBD.DstRGB);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.OneMinusSrcAAA, OpAC.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedInvertOvgPremul(ref UcodeAssembler asm)
+        {
+            asm.Sub(CC.T, Dest.PBR, OpBD.ConstantOne, OpBD.DstRGB);
+            asm.Mmadd(CC.T, Dest.Temp0, OpAC.SrcAAA, OpBD.PBR, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedLinearDodgePremul(ref UcodeAssembler asm)
+        {
+            asm.Mmadd(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedLinearBurnPremul(ref UcodeAssembler asm)
+        {
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedVividLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantRGB);
+            asm.Sub(CC.GE, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+            asm.Add(CC.GE, Dest.PBR, OpBD.PBR, OpBD.PBR);
+            asm.Rcp(CC.GE, Dest.PBR, OpAC.PBR);
+            asm.Mul(CC.GE, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GE, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Add(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Rcp(CC.LT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.LT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantZero);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantOne);
+            asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedLinearLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 2f, 2f, 2f);
+            asm.Madd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedPinLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.Temp0, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+            asm.Add(CC.LE, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Min(CC.LE, Dest.Temp0, OpAC.PBR, OpBD.Temp1);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedHardMixPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Mul(CC.LT, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedRedPremul(ref UcodeAssembler asm)
+        {
+            asm.Mov(CC.T, Dest.Temp0, OpBD.DstRGB);
+            asm.Mov(CC.T, Dest.Temp0.R, OpBD.SrcRGB);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedGreenPremul(ref UcodeAssembler asm)
+        {
+            asm.Mov(CC.T, Dest.Temp0, OpBD.DstRGB);
+            asm.Mov(CC.T, Dest.Temp0.G, OpBD.SrcRGB);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedBluePremul(ref UcodeAssembler asm)
+        {
+            asm.Mov(CC.T, Dest.Temp0, OpBD.DstRGB);
+            asm.Mov(CC.T, Dest.Temp0.B, OpBD.SrcRGB);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedHslHuePremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+            asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp2, OpAC.Temp0, OpBD.PBR);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.Temp2.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp2);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedHslSaturationPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+            asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp1, OpAC.Temp0, OpBD.PBR);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.GT, Dest.Temp1.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp1);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedHslColorPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp2, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp2, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedHslLuminosityPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp2, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp2.BBB, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp1, OpBD.Temp1, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp2);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp1, OpBD.Temp2);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp2);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp1, OpBD.Temp2, OpAC.Temp2, OpBD.Temp2);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp2);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.SrcRGB, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenDisjointSrcPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+        }
+
+        private static FixedFunctionAlpha GenDisjointDstPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.Temp1, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenDisjointSrcOverPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp2);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointDstOverPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp1);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointSrcInPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Sub(CC.T, Dest.Temp1.RToA, OpBD.DstAAA, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointDstInPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.Temp1, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Sub(CC.T, Dest.Temp1.RToA, OpBD.DstAAA, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointSrcOutPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointDstOutPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointSrcAtopPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenDisjointDstAtopPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.Temp1, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+        }
+
+        private static FixedFunctionAlpha GenDisjointXorPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            asm.Min(CC.T, Dest.Temp1, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Add(CC.T, Dest.Temp1.RToA, OpBD.Temp1, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointPlusPremul(ref UcodeAssembler asm)
+        {
+            asm.Add(CC.T, Dest.Temp0, OpBD.DstRGB, OpBD.SrcRGB);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenDisjointMultiplyPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointScreenPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointOverlayPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantRGB);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp2, OpBD.Temp1, OpAC.Temp2, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointDarkenPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointLightenPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Max(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointColorDodgePremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.ConstantOne, OpBD.Temp2);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.Temp0);
+            asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp1, OpBD.ConstantZero);
+            asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantZero);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointColorBurnPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.Temp2, OpBD.ConstantZero);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.Temp2);
+            asm.Mmsub(CC.GT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Max(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointHardLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.Temp2, OpBD.ConstantRGB);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp2, OpBD.Temp1, OpAC.Temp2, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointSoftLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(4, 0.25f, 0.25f, 0.25f);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantRGB);
+            asm.SetConstant(0, 0.2605f, 0.2605f, 0.2605f);
+            asm.Mul(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(1, -0.7817f, -0.7817f, -0.7817f);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(2, 0.3022f, 0.3022f, 0.3022f);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(3, 0.2192f, 0.2192f, 0.2192f);
+            asm.Add(CC.GT, Dest.Temp0, OpBD.PBR, OpBD.ConstantRGB);
+            asm.SetConstant(5, 16f, 16f, 16f);
+            asm.Mul(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(6, 12f, 12f, 12f);
+            asm.Mmsub(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(7, 3f, 3f, 3f);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Mmsub(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.ConstantOne, OpAC.Temp1, OpBD.Temp1);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointDifferencePremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.Temp0, OpBD.Temp2, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointExclusionPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.Temp1);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointInvertPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp0, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenDisjointInvertRGBPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.ConstantOne, OpAC.Temp2, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp0, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenDisjointLinearDodgePremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointLinearBurnPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointVividLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantRGB);
+            asm.Sub(CC.GE, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+            asm.Add(CC.GE, Dest.PBR, OpBD.PBR, OpBD.PBR);
+            asm.Rcp(CC.GE, Dest.PBR, OpAC.PBR);
+            asm.Mul(CC.GE, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GE, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Add(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Rcp(CC.LT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.LT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantZero);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantOne);
+            asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointLinearLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 2f, 2f, 2f);
+            asm.Madd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointPinLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.Temp0, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+            asm.Add(CC.LE, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Min(CC.LE, Dest.Temp0, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointHardMixPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Mul(CC.LT, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointHslHuePremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+            asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp2, OpAC.Temp0, OpBD.PBR);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.Temp2.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp2);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointHslSaturationPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+            asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp1, OpAC.Temp0, OpBD.PBR);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.GT, Dest.Temp1.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp1);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointHslColorPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp2, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp2, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointHslLuminosityPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp2, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp2.BBB, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp1, OpBD.Temp1, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp2);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp1, OpBD.Temp2);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp2);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp1, OpBD.Temp2, OpAC.Temp2, OpBD.Temp2);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp2);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.Temp2, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenConjointSrcPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointDstPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointSrcOverPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp2, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointDstOverPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointSrcInPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MinimumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointDstInPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MinimumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointSrcOutPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Max(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantZero);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenConjointDstOutPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Max(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantZero);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenConjointSrcAtopPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointDstAtopPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointXorPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            asm.Sub(CC.T, Dest.Temp1.CC, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Sub(CC.LT, Dest.Temp1, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mov(CC.T, Dest.Temp1.RToA, OpBD.Temp1);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenConjointMultiplyPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointScreenPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointOverlayPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantRGB);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp2, OpBD.Temp1, OpAC.Temp2, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointDarkenPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointLightenPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Max(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointColorDodgePremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.ConstantOne, OpBD.Temp2);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.Temp0);
+            asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp1, OpBD.ConstantZero);
+            asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointColorBurnPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.Temp2, OpBD.ConstantZero);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.Temp2);
+            asm.Mmsub(CC.GT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Max(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointHardLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.Temp2, OpBD.ConstantRGB);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp2, OpBD.Temp1, OpAC.Temp2, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointSoftLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(4, 0.25f, 0.25f, 0.25f);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantRGB);
+            asm.SetConstant(0, 0.2605f, 0.2605f, 0.2605f);
+            asm.Mul(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(1, -0.7817f, -0.7817f, -0.7817f);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(2, 0.3022f, 0.3022f, 0.3022f);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(3, 0.2192f, 0.2192f, 0.2192f);
+            asm.Add(CC.GT, Dest.Temp0, OpBD.PBR, OpBD.ConstantRGB);
+            asm.SetConstant(5, 16f, 16f, 16f);
+            asm.Mul(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(6, 12f, 12f, 12f);
+            asm.Mmsub(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(7, 3f, 3f, 3f);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Mmsub(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.ConstantOne, OpAC.Temp1, OpBD.Temp1);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointDifferencePremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.Temp0, OpBD.Temp2, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointExclusionPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.Temp1);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointInvertPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointInvertRGBPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.ConstantOne, OpAC.Temp2, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointLinearDodgePremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointLinearBurnPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointVividLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantRGB);
+            asm.Sub(CC.GE, Dest.PBR, OpBD.ConstantOne, OpBD.Temp2);
+            asm.Add(CC.GE, Dest.PBR, OpBD.PBR, OpBD.PBR);
+            asm.Rcp(CC.GE, Dest.PBR, OpAC.PBR);
+            asm.Mul(CC.GE, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GE, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Add(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Rcp(CC.LT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.LT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantZero);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantOne);
+            asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointLinearLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 2f, 2f, 2f);
+            asm.Madd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointPinLightPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.Temp0, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+            asm.Add(CC.LE, Dest.PBR, OpBD.Temp2, OpBD.Temp2);
+            asm.Min(CC.LE, Dest.Temp0, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointHardMixPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Mul(CC.LT, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointHslHuePremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+            asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp2, OpAC.Temp0, OpBD.PBR);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.Temp2.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp2);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointHslSaturationPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+            asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp1, OpAC.Temp0, OpBD.PBR);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.GT, Dest.Temp1.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp1);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointHslColorPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp2, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp2, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointHslLuminosityPremul(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp2, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp2.BBB, OpAC.Temp2, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp1, OpBD.Temp1, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp2);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp1, OpBD.Temp2);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp2);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp1, OpBD.Temp2, OpAC.Temp2, OpBD.Temp2);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp2);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.PBR);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp2, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedDstOver(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.SrcAAA);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedSrcIn(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.PBR, OpBD.DstAAA);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.DstAlphaGl, BlendFactor.ZeroGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedSrcOut(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.PBR, OpBD.OneMinusDstAAA);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneMinusDstAlphaGl, BlendFactor.ZeroGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedSrcAtop(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.PBR, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.OneMinusSrcAAA, OpAC.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedDstAtop(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.SrcAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedXor(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.PBR, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.OneMinusSrcAAA, OpAC.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneMinusDstAlphaGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedPlusClamped(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Add(CC.T, Dest.PBR, OpBD.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedPlusClampedAlpha(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+            asm.Min(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedPlusDarker(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+            asm.Add(CC.T, Dest.PBR, OpBD.PBR, OpBD.Temp2);
+            asm.Add(CC.T, Dest.PBR, OpBD.PBR, OpBD.DstRGB);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.SrcAAA);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.DstAAA);
+            asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedMultiply(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.PBR, OpBD.DstRGB);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedScreen(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.PBR, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.DstRGB);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedOverlay(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantRGB);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.SrcRGB, OpBD.Temp1, OpAC.SrcRGB, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedDarken(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.PBR, OpBD.DstAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.SrcAAA);
+            asm.Min(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedLighten(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.PBR, OpBD.DstAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.SrcAAA);
+            asm.Max(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedColorDodge(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.PBR);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.SrcAAA);
+            asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.DstRGB);
+            asm.Min(CC.GT, Dest.PBR, OpAC.DstAAA, OpBD.PBR);
+            asm.Mul(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.SrcAAA);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.DstRGB, OpBD.ConstantZero);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedColorBurn(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.DstAAA, OpBD.SrcAAA, OpAC.SrcAAA, OpBD.DstRGB);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.Temp2);
+            asm.Mul(CC.T, Dest.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.PBR);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.SrcAAA, OpBD.DstAAA, OpAC.SrcAAA, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp2, OpBD.ConstantZero);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.DstAAA, OpBD.DstRGB);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedHardLight(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.SrcRGB, OpBD.ConstantRGB);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.SrcRGB, OpBD.Temp1, OpAC.SrcRGB, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedSoftLight(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(4, 0.25f, 0.25f, 0.25f);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantRGB);
+            asm.SetConstant(0, 0.2605f, 0.2605f, 0.2605f);
+            asm.Mul(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(1, -0.7817f, -0.7817f, -0.7817f);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(2, 0.3022f, 0.3022f, 0.3022f);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(3, 0.2192f, 0.2192f, 0.2192f);
+            asm.Add(CC.GT, Dest.Temp0, OpBD.PBR, OpBD.ConstantRGB);
+            asm.SetConstant(5, 16f, 16f, 16f);
+            asm.Mul(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(6, 12f, 12f, 12f);
+            asm.Mmsub(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(7, 3f, 3f, 3f);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Mmsub(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.ConstantOne, OpAC.Temp1, OpBD.Temp1);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedDifference(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.SrcRGB);
+            asm.Sub(CC.LT, Dest.Temp0, OpBD.SrcRGB, OpBD.Temp1);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedMinus(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Sub(CC.T, Dest.Temp0, OpBD.DstRGB, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.ReverseSubtractGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedMinusClamped(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Sub(CC.T, Dest.PBR, OpBD.DstRGB, OpBD.PBR);
+            asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Max(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantZero);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedExclusion(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.PBR, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.DstRGB);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.Temp2, OpBD.DstRGB);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedContrast(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.SetConstant(0, 2f, 2f, 2f);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.ConstantRGB, OpAC.DstAAA, OpBD.ConstantOne);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.Temp2, OpBD.ConstantRGB, OpAC.SrcAAA, OpBD.ConstantOne);
+            asm.Mul(CC.T, Dest.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.PBR, OpBD.DstAAA);
+            asm.SetConstant(1, 0.5f, 0.5f, 0.5f);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantRGB);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedInvertRGB(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.DstAAA, OpAC.PBR, OpBD.DstRGB);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.DstRGB, OpBD.OneMinusSrcAAA, OpAC.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedLinearDodge(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mmadd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedLinearBurn(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.PBR, OpBD.DstAAA, OpAC.DstRGB, OpBD.SrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+            asm.Mmadd(CC.T, Dest.PBR, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.Temp0, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedVividLight(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantRGB);
+            asm.Sub(CC.GE, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+            asm.Add(CC.GE, Dest.PBR, OpBD.PBR, OpBD.PBR);
+            asm.Rcp(CC.GE, Dest.PBR, OpAC.PBR);
+            asm.Mul(CC.GE, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GE, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Add(CC.LT, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Rcp(CC.LT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.LT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantZero);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantOne);
+            asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedLinearLight(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 2f, 2f, 2f);
+            asm.Madd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedPinLight(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Sub(CC.T, Dest.Temp0, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+            asm.Add(CC.LE, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Min(CC.LE, Dest.Temp0, OpAC.PBR, OpBD.Temp1);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedHardMix(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Mul(CC.LT, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.Temp2, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedRed(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.DstRGB);
+            asm.Mov(CC.T, Dest.Temp0.R, OpBD.Temp2);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedGreen(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.DstRGB);
+            asm.Mov(CC.T, Dest.Temp0.G, OpBD.Temp2);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedBlue(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.Temp2, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.DstRGB);
+            asm.Mov(CC.T, Dest.Temp0.B, OpBD.Temp2);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedHslHue(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+            asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.SrcRGB, OpAC.Temp0, OpBD.PBR);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.Temp2.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp2);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.PBR, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedHslSaturation(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+            asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp1, OpAC.Temp0, OpBD.PBR);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Min(CC.GT, Dest.Temp1.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp1);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.PBR, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedHslColor(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.SrcRGB, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.SrcRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.PBR, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenUncorrelatedHslLuminosity(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.SrcRGB, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp2.BBB, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp1, OpBD.Temp1, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp2);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp1, OpBD.Temp2);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp2);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp1, OpBD.Temp2, OpAC.Temp2, OpBD.Temp2);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp2);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Mmadd(CC.T, Dest.Temp1, OpAC.PBR, OpBD.OneMinusDstAAA, OpAC.DstRGB, OpBD.OneMinusSrcAAA);
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.DstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl);
+        }
+
+        private static FixedFunctionAlpha GenDisjointSrc(ref UcodeAssembler asm)
+        {
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+        }
+
+        private static FixedFunctionAlpha GenDisjointSrcOver(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.SrcRGB);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointDstOver(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp1);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointSrcIn(ref UcodeAssembler asm)
+        {
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Sub(CC.T, Dest.Temp1.RToA, OpBD.DstAAA, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointSrcOut(ref UcodeAssembler asm)
+        {
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointSrcAtop(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenDisjointDstAtop(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.Temp1, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+        }
+
+        private static FixedFunctionAlpha GenDisjointXor(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            asm.Min(CC.T, Dest.Temp1, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Add(CC.T, Dest.Temp1.RToA, OpBD.Temp1, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointPlus(ref UcodeAssembler asm)
+        {
+            asm.Mul(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.SrcAAA);
+            asm.Add(CC.T, Dest.Temp0, OpBD.DstRGB, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenDisjointMultiply(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointScreen(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointOverlay(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantRGB);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.SrcRGB, OpBD.Temp1, OpAC.SrcRGB, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointDarken(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointLighten(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Max(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointColorDodge(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.ConstantOne, OpBD.SrcRGB);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.Temp0);
+            asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp1, OpBD.ConstantZero);
+            asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantZero);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointColorBurn(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.SrcRGB, OpBD.ConstantZero);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.SrcRGB);
+            asm.Mmsub(CC.GT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Max(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointHardLight(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.SrcRGB, OpBD.ConstantRGB);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.SrcRGB, OpBD.Temp1, OpAC.SrcRGB, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointSoftLight(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(4, 0.25f, 0.25f, 0.25f);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantRGB);
+            asm.SetConstant(0, 0.2605f, 0.2605f, 0.2605f);
+            asm.Mul(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(1, -0.7817f, -0.7817f, -0.7817f);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(2, 0.3022f, 0.3022f, 0.3022f);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(3, 0.2192f, 0.2192f, 0.2192f);
+            asm.Add(CC.GT, Dest.Temp0, OpBD.PBR, OpBD.ConstantRGB);
+            asm.SetConstant(5, 16f, 16f, 16f);
+            asm.Mul(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(6, 12f, 12f, 12f);
+            asm.Mmsub(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(7, 3f, 3f, 3f);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Mmsub(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.ConstantOne, OpAC.Temp1, OpBD.Temp1);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointDifference(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.SrcRGB);
+            asm.Sub(CC.LT, Dest.Temp0, OpBD.SrcRGB, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointExclusion(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.Temp1);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointInvertRGB(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.Temp0, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenDisjointLinearDodge(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointLinearBurn(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointVividLight(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantRGB);
+            asm.Sub(CC.GE, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+            asm.Add(CC.GE, Dest.PBR, OpBD.PBR, OpBD.PBR);
+            asm.Rcp(CC.GE, Dest.PBR, OpAC.PBR);
+            asm.Mul(CC.GE, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GE, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Add(CC.LT, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Rcp(CC.LT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.LT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantZero);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantOne);
+            asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointLinearLight(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 2f, 2f, 2f);
+            asm.Madd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointPinLight(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Sub(CC.T, Dest.Temp0, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+            asm.Add(CC.LE, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Min(CC.LE, Dest.Temp0, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointHardMix(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Mul(CC.LT, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointHslHue(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+            asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.SrcRGB, OpAC.Temp0, OpBD.PBR);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.Temp2.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp2);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointHslSaturation(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+            asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp1, OpAC.Temp0, OpBD.PBR);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Min(CC.GT, Dest.Temp1.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp1);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointHslColor(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.SrcRGB, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.SrcRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenDisjointHslLuminosity(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.SrcRGB, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp2.BBB, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp1, OpBD.Temp1, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp2);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp1, OpBD.Temp2);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp2);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp1, OpBD.Temp2, OpAC.Temp2, OpBD.Temp2);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp2);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.OneMinusSrcAAA);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.Temp1, OpAC.PBR, OpBD.Temp0);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.SrcAAA, OpBD.OneMinusDstAAA);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.PBR, OpBD.SrcRGB, OpAC.Temp0);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Min(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenConjointSrc(ref UcodeAssembler asm)
+        {
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointSrcOver(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.SrcRGB);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.SrcRGB, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointDstOver(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp1, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointSrcIn(ref UcodeAssembler asm)
+        {
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MinimumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointSrcOut(ref UcodeAssembler asm)
+        {
+            asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Max(CC.T, Dest.Temp1.RToA, OpAC.PBR, OpBD.ConstantZero);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenConjointSrcAtop(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointDstAtop(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointXor(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            asm.Sub(CC.T, Dest.Temp1.CC, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Sub(CC.LT, Dest.Temp1, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mov(CC.T, Dest.Temp1.RToA, OpBD.Temp1);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.Temp0);
+            return FixedFunctionAlpha.Disabled;
+        }
+
+        private static FixedFunctionAlpha GenConjointMultiply(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointScreen(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointOverlay(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantRGB);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.SrcRGB, OpBD.Temp1, OpAC.SrcRGB, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointDarken(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointLighten(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Max(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointColorDodge(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.ConstantOne, OpBD.SrcRGB);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.Temp0);
+            asm.Mul(CC.GT, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.Temp1, OpBD.ConstantZero);
+            asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointColorBurn(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.SrcRGB, OpBD.ConstantZero);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.SrcRGB);
+            asm.Mmsub(CC.GT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Max(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Mov(CC.LE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointHardLight(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.SrcRGB, OpBD.ConstantRGB);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.SrcRGB, OpBD.Temp1, OpAC.SrcRGB, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.GT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointSoftLight(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(4, 0.25f, 0.25f, 0.25f);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantRGB);
+            asm.SetConstant(0, 0.2605f, 0.2605f, 0.2605f);
+            asm.Mul(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(1, -0.7817f, -0.7817f, -0.7817f);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(2, 0.3022f, 0.3022f, 0.3022f);
+            asm.Mmadd(CC.GT, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(3, 0.2192f, 0.2192f, 0.2192f);
+            asm.Add(CC.GT, Dest.Temp0, OpBD.PBR, OpBD.ConstantRGB);
+            asm.SetConstant(5, 16f, 16f, 16f);
+            asm.Mul(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(6, 12f, 12f, 12f);
+            asm.Mmsub(CC.LE, Dest.PBR, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.SetConstant(7, 3f, 3f, 3f);
+            asm.Mmadd(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Mmsub(CC.LE, Dest.Temp0, OpAC.Temp1, OpBD.ConstantOne, OpAC.Temp1, OpBD.Temp1);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointDifference(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.SrcRGB);
+            asm.Sub(CC.LT, Dest.Temp0, OpBD.SrcRGB, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointExclusion(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+            asm.Mmsub(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.Temp1);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointInvertRGB(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mmsub(CC.T, Dest.Temp0, OpAC.SrcRGB, OpBD.ConstantOne, OpAC.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR, OpAC.DstAAA, OpBD.SrcAAA);
+            asm.Mul(CC.T, Dest.Temp0, OpAC.Temp0, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Madd(CC.T, Dest.Temp0, OpAC.Temp1, OpBD.PBR, OpAC.Temp0);
+            return new FixedFunctionAlpha(BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointLinearDodge(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+            asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointLinearBurn(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Max(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointVividLight(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.5f, 0.5f, 0.5f);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantRGB);
+            asm.Sub(CC.GE, Dest.PBR, OpBD.ConstantOne, OpBD.SrcRGB);
+            asm.Add(CC.GE, Dest.PBR, OpBD.PBR, OpBD.PBR);
+            asm.Rcp(CC.GE, Dest.PBR, OpAC.PBR);
+            asm.Mul(CC.GE, Dest.PBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GE, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Add(CC.LT, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Rcp(CC.LT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.LT, Dest.PBR, OpAC.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.LT, Dest.Temp0, OpBD.ConstantOne, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantZero);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcRGB, OpBD.ConstantOne);
+            asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointLinearLight(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 2f, 2f, 2f);
+            asm.Madd(CC.T, Dest.PBR, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Max(CC.T, Dest.PBR, OpAC.PBR, OpBD.ConstantZero);
+            asm.Min(CC.T, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointPinLight(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Sub(CC.T, Dest.Temp0, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.ConstantZero);
+            asm.Add(CC.LE, Dest.PBR, OpBD.SrcRGB, OpBD.SrcRGB);
+            asm.Min(CC.LE, Dest.Temp0, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointHardMix(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Add(CC.T, Dest.PBR, OpBD.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Mul(CC.LT, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Mov(CC.GE, Dest.Temp0, OpBD.ConstantOne);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointHslHue(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+            asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.SrcRGB, OpAC.Temp0, OpBD.PBR);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.Temp2.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp2);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointHslSaturation(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.PBR);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.Temp0.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.Temp0.CC, OpBD.PBR, OpBD.Temp0);
+            asm.Rcp(CC.GT, Dest.Temp0, OpAC.Temp0);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.Temp1, OpAC.Temp0, OpBD.PBR);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Min(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Min(CC.GT, Dest.Temp1.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Mov(CC.GT, Dest.PBR.GBR, OpBD.SrcRGB);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Max(CC.GT, Dest.PBR.GBR, OpAC.PBR, OpBD.SrcRGB);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp0, OpBD.Temp1);
+            asm.Mul(CC.LE, Dest.Temp0, OpAC.SrcAAA, OpBD.ConstantZero);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp0, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp0, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointHslColor(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.PBR, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp1.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.SrcRGB, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp2, OpBD.SrcRGB, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp1);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp2, OpBD.Temp1);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp1);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp2);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp1, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp2, OpBD.Temp1, OpAC.Temp1, OpBD.Temp1);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp1);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+
+        private static FixedFunctionAlpha GenConjointHslLuminosity(ref UcodeAssembler asm)
+        {
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.SetConstant(0, 0.3f, 0.59f, 0.11f);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.SrcRGB, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.Temp2.BBB, OpAC.SrcRGB, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Mul(CC.T, Dest.PBR.RRR, OpAC.Temp1, OpBD.ConstantRGB);
+            asm.Madd(CC.T, Dest.PBR.GGG, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Madd(CC.T, Dest.PBR.BBB, OpAC.Temp1, OpBD.ConstantRGB, OpAC.PBR);
+            asm.Sub(CC.T, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Add(CC.T, Dest.Temp1, OpBD.Temp1, OpBD.PBR);
+            asm.Mov(CC.T, Dest.Temp0, OpBD.PBR);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Max(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.PBR, OpBD.ConstantOne);
+            asm.Add(CC.GT, Dest.PBR, OpBD.PBR, OpBD.ConstantOne);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.PBR, OpBD.Temp2);
+            asm.Rcp(CC.GT, Dest.PBR, OpAC.PBR);
+            asm.Mmsub(CC.GT, Dest.Temp0, OpAC.PBR, OpBD.ConstantOne, OpAC.PBR, OpBD.Temp2);
+            asm.Sub(CC.GT, Dest.PBR, OpBD.Temp1, OpBD.Temp2);
+            asm.Madd(CC.GT, Dest.Temp0, OpAC.Temp0, OpBD.PBR, OpAC.Temp2);
+            asm.Mov(CC.T, Dest.PBR.GBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR.GBR, OpAC.PBR, OpBD.Temp1);
+            asm.Min(CC.T, Dest.PBR.GBR.CC, OpAC.PBR, OpBD.Temp1);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.Temp2, OpBD.PBR);
+            asm.Rcp(CC.LT, Dest.Temp0, OpAC.PBR);
+            asm.Mmsub(CC.LT, Dest.PBR, OpAC.Temp1, OpBD.Temp2, OpAC.Temp2, OpBD.Temp2);
+            asm.Madd(CC.LT, Dest.Temp0, OpAC.PBR, OpBD.Temp0, OpAC.Temp2);
+            asm.Rcp(CC.T, Dest.PBR, OpAC.DstAAA);
+            asm.Mul(CC.T, Dest.Temp1, OpAC.DstRGB, OpBD.PBR);
+            asm.Sub(CC.T, Dest.PBR.CC, OpBD.SrcAAA, OpBD.DstAAA);
+            asm.Mmadd(CC.GE, Dest.Temp0, OpAC.Temp0, OpBD.DstAAA, OpAC.SrcRGB, OpBD.PBR);
+            asm.Sub(CC.LT, Dest.PBR, OpBD.DstAAA, OpBD.SrcAAA);
+            asm.Mmadd(CC.LT, Dest.Temp0, OpAC.Temp0, OpBD.SrcAAA, OpAC.Temp1, OpBD.PBR);
+            return new FixedFunctionAlpha(BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl);
+        }
+    }
+}
\ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendManager.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendManager.cs
new file mode 100644
index 00000000..8072c6af
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendManager.cs
@@ -0,0 +1,115 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.GAL;
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed.Blender
+{
+    /// <summary>
+    /// Advanced blend manager.
+    /// </summary>
+    class AdvancedBlendManager
+    {
+        private const int InstructionRamSize = 128;
+        private const int InstructionRamSizeMask = InstructionRamSize - 1;
+
+        private readonly DeviceStateWithShadow<ThreedClassState> _state;
+
+        private readonly uint[] _code;
+        private int _ip;
+
+        /// <summary>
+        /// Creates a new instance of the advanced blend manager.
+        /// </summary>
+        /// <param name="state">GPU state of the channel owning this manager</param>
+        public AdvancedBlendManager(DeviceStateWithShadow<ThreedClassState> state)
+        {
+            _state = state;
+            _code = new uint[InstructionRamSize];
+        }
+
+        /// <summary>
+        /// Sets the start offset of the blend microcode in memory.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void LoadBlendUcodeStart(int argument)
+        {
+            _ip = argument;
+        }
+
+        /// <summary>
+        /// Pushes one word of blend microcode.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        public void LoadBlendUcodeInstruction(int argument)
+        {
+            _code[_ip++ & InstructionRamSizeMask] = (uint)argument;
+        }
+
+        /// <summary>
+        /// Tries to identify the current advanced blend function being used,
+        /// given the current state and microcode that was uploaded.
+        /// </summary>
+        /// <param name="descriptor">Advanced blend descriptor</param>
+        /// <returns>True if the function was found, false otherwise</returns>
+        public bool TryGetAdvancedBlend(out AdvancedBlendDescriptor descriptor)
+        {
+            Span<uint> currentCode = new Span<uint>(_code);
+            byte codeLength = (byte)_state.State.BlendUcodeSize;
+
+            if (currentCode.Length > codeLength)
+            {
+                currentCode = currentCode.Slice(0, codeLength);
+            }
+
+            Hash128 hash = XXHash128.ComputeHash(MemoryMarshal.Cast<uint, byte>(currentCode));
+
+            descriptor = default;
+
+            if (!AdvancedBlendPreGenTable.Entries.TryGetValue(hash, out var entry))
+            {
+                return false;
+            }
+
+            if (entry.Constants != null)
+            {
+                bool constantsMatch = true;
+
+                for (int i = 0; i < entry.Constants.Length; i++)
+                {
+                    RgbFloat constant = entry.Constants[i];
+                    RgbHalf constant2 = _state.State.BlendUcodeConstants[i];
+
+                    if ((Half)constant.R != constant2.UnpackR() ||
+                        (Half)constant.G != constant2.UnpackG() ||
+                        (Half)constant.B != constant2.UnpackB())
+                    {
+                        constantsMatch = false;
+                        break;
+                    }
+                }
+
+                if (!constantsMatch)
+                {
+                    return false;
+                }
+            }
+
+            if (entry.Alpha.Enable != _state.State.BlendUcodeEnable)
+            {
+                return false;
+            }
+
+            if (entry.Alpha.Enable == BlendUcodeEnable.EnableRGBA &&
+                (entry.Alpha.AlphaOp != _state.State.BlendStateCommon.AlphaOp ||
+                entry.Alpha.AlphaSrcFactor != _state.State.BlendStateCommon.AlphaSrcFactor ||
+                entry.Alpha.AlphaDstFactor != _state.State.BlendStateCommon.AlphaDstFactor))
+            {
+                return false;
+            }
+
+            descriptor = new AdvancedBlendDescriptor(entry.Op, entry.Overlap, entry.SrcPreMultiplied);
+            return true;
+        }
+    }
+}
\ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendPreGenTable.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendPreGenTable.cs
new file mode 100644
index 00000000..d35d8abf
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendPreGenTable.cs
@@ -0,0 +1,273 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.GAL;
+using System;
+using System.Collections.Generic;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed.Blender
+{
+    /// <summary>
+    /// Advanced blend function entry.
+    /// </summary>
+    struct AdvancedBlendEntry
+    {
+        /// <summary>
+        /// Advanced blend operation.
+        /// </summary>
+        public AdvancedBlendOp Op { get; }
+
+        /// <summary>
+        /// Advanced blend overlap mode.
+        /// </summary>
+        public AdvancedBlendOverlap Overlap { get; }
+
+        /// <summary>
+        /// Whenever the source input is pre-multiplied.
+        /// </summary>
+        public bool SrcPreMultiplied { get; }
+
+        /// <summary>
+        /// Constants used by the microcode.
+        /// </summary>
+        public RgbFloat[] Constants { get; }
+
+        /// <summary>
+        /// Fixed function alpha state.
+        /// </summary>
+        public FixedFunctionAlpha Alpha { get; }
+
+        /// <summary>
+        /// Creates a new advanced blend function entry.
+        /// </summary>
+        /// <param name="op">Advanced blend operation</param>
+        /// <param name="overlap">Advanced blend overlap mode</param>
+        /// <param name="srcPreMultiplied">Whenever the source input is pre-multiplied</param>
+        /// <param name="constants">Constants used by the microcode</param>
+        /// <param name="alpha">Fixed function alpha state</param>
+        public AdvancedBlendEntry(
+            AdvancedBlendOp op,
+            AdvancedBlendOverlap overlap,
+            bool srcPreMultiplied,
+            RgbFloat[] constants,
+            FixedFunctionAlpha alpha)
+        {
+            Op = op;
+            Overlap = overlap;
+            SrcPreMultiplied = srcPreMultiplied;
+            Constants = constants;
+            Alpha = alpha;
+        }
+    }
+
+    /// <summary>
+    /// Pre-generated hash table with advanced blend functions used by the driver.
+    /// </summary>
+    static class AdvancedBlendPreGenTable
+    {
+        /// <summary>
+        /// Advanced blend functions dictionary.
+        /// </summary>
+        public static readonly IReadOnlyDictionary<Hash128, AdvancedBlendEntry> Entries = new Dictionary<Hash128, AdvancedBlendEntry>()
+        {
+            { new Hash128(0x19ECF57B83DE31F7, 0x5BAE759246F264C0), new AdvancedBlendEntry(AdvancedBlendOp.PlusClamped, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xDE1B14A356A1A9ED, 0x59D803593C607C1D), new AdvancedBlendEntry(AdvancedBlendOp.PlusClampedAlpha, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x1A3C3A6D32DEC368, 0xBCAE519EC6AAA045), new AdvancedBlendEntry(AdvancedBlendOp.PlusDarker, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x6FD380261A63B240, 0x17C3B335DBB9E3DB), new AdvancedBlendEntry(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x1D39164823D3A2D1, 0xC45350959CE1C8FB), new AdvancedBlendEntry(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x18DF09FF53B129FE, 0xC02EDA33C36019F6), new AdvancedBlendEntry(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x5973E583271EBF06, 0x711497D75D1272E0), new AdvancedBlendEntry(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x4759E0E5DA54D5E8, 0x1FDD57C0C38AFA1F), new AdvancedBlendEntry(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x337684D43CCE97FA, 0x0139E30CC529E1C9), new AdvancedBlendEntry(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0xDA59E85D8428992D, 0x1D3D7C64C9EF0132), new AdvancedBlendEntry(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x9455B949298CE805, 0xE73D3301518BE98A), new AdvancedBlendEntry(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0xBDD3B4DEDBE336AA, 0xBFA4DCD50D535DEE), new AdvancedBlendEntry(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.2605f, 0.2605f, 0.2605f), new RgbFloat(-0.7817f, -0.7817f, -0.7817f), new RgbFloat(0.3022f, 0.3022f, 0.3022f), new RgbFloat(0.2192f, 0.2192f, 0.2192f), new RgbFloat(0.25f, 0.25f, 0.25f), new RgbFloat(16f, 16f, 16f), new RgbFloat(12f, 12f, 12f), new RgbFloat(3f, 3f, 3f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x22D4E970A028649A, 0x4F3FCB055FCED965), new AdvancedBlendEntry(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0xA346A91311D72114, 0x151A27A3FB0A1904), new AdvancedBlendEntry(AdvancedBlendOp.Minus, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.ReverseSubtractGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x8A307241061FACD6, 0xA39D1826440B8EE7), new AdvancedBlendEntry(AdvancedBlendOp.MinusClamped, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xB3BE569485EFFFE0, 0x0BA4E269B3CFB165), new AdvancedBlendEntry(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x36FCA3277DC11822, 0x2BC0F6CAC2029672), new AdvancedBlendEntry(AdvancedBlendOp.Contrast, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(2f, 2f, 2f), new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x4A6226AF2DE9BD7F, 0xEB890D7DA716F73A), new AdvancedBlendEntry(AdvancedBlendOp.Invert, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0xF364CAA94E160FEB, 0xBF364512C72A3797), new AdvancedBlendEntry(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x6BF791AB4AC19C87, 0x6FA17A994EA0FCDE), new AdvancedBlendEntry(AdvancedBlendOp.InvertOvg, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x053C75A0AE0BB222, 0x03C791FEEB59754C), new AdvancedBlendEntry(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x25762AB40B6CBDE9, 0x595E9A968AC4F01C), new AdvancedBlendEntry(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0xC2D05E2DBE16955D, 0xB8659C7A3FCFA7CE), new AdvancedBlendEntry(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x223F220B8F74CBFB, 0xD3DD19D7C39209A5), new AdvancedBlendEntry(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(2f, 2f, 2f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0xD0DAE57A9F1FE78A, 0x353796BCFB8CE30B), new AdvancedBlendEntry(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x601C8CBEC07FF8FF, 0xB8E22882360E8695), new AdvancedBlendEntry(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x3A55B7B78C76A7A8, 0x206F503B2D9FFEAA), new AdvancedBlendEntry(AdvancedBlendOp.Red, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x80BC65C7831388E5, 0xC652457B2C766AEC), new AdvancedBlendEntry(AdvancedBlendOp.Green, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x3D3A912E5833EE13, 0x307895951349EE33), new AdvancedBlendEntry(AdvancedBlendOp.Blue, AdvancedBlendOverlap.Uncorrelated, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x289105BE92E81803, 0xFD8F1F03D15C53B4), new AdvancedBlendEntry(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x007AE3BD140764EB, 0x0EE05A0D2E80BBAE), new AdvancedBlendEntry(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x77F7EE0DB3FDDB96, 0xDEA47C881306DB3E), new AdvancedBlendEntry(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x66F4E9A7D73CA157, 0x1486058A177DB11C), new AdvancedBlendEntry(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Uncorrelated, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x593E9F331612D618, 0x9D217BEFA4EB919A), new AdvancedBlendEntry(AdvancedBlendOp.Src, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+            { new Hash128(0x0A5194C5E6891106, 0xDD8EC6586106557C), new AdvancedBlendEntry(AdvancedBlendOp.Dst, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x8D77173D5E06E916, 0x06AB190E7D10F4D4), new AdvancedBlendEntry(AdvancedBlendOp.SrcOver, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x655B4EBC148981DA, 0x455999EF2B9BD28A), new AdvancedBlendEntry(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x98F5437D5F518929, 0xBFF4A6E83183DB63), new AdvancedBlendEntry(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x6ADDEFE3B9CEF2FD, 0xB6F6272AFECB1AAB), new AdvancedBlendEntry(AdvancedBlendOp.DstIn, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x80953F0953BF05B1, 0xD59ABFAA34F8196F), new AdvancedBlendEntry(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xA401D9AA2A39C121, 0xFC0C8005C22AD7E3), new AdvancedBlendEntry(AdvancedBlendOp.DstOut, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x06274FB7CA9CDD22, 0x6CE8188B1A9AB6EF), new AdvancedBlendEntry(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x0B079BE7F7F70817, 0xB72E7736CA51E321), new AdvancedBlendEntry(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+            { new Hash128(0x66215C99403CEDDE, 0x900B733D62204C48), new AdvancedBlendEntry(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x12DEF2AD900CAD6C, 0x58CF5CC3004910DF), new AdvancedBlendEntry(AdvancedBlendOp.Plus, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x272BA3A49F64DAE4, 0xAC70B96C00A99EAF), new AdvancedBlendEntry(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x206C34AAA7D3F545, 0xDA4B30CACAA483A0), new AdvancedBlendEntry(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x3D93494920D257BE, 0xDCC573BE1F5F4449), new AdvancedBlendEntry(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x0D7417D80191107B, 0xEAF40547827E005F), new AdvancedBlendEntry(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xEC1B03E8C883F9C9, 0x2D3CA044C58C01B4), new AdvancedBlendEntry(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x58A19A0135D68B31, 0x82F35B97AED068E5), new AdvancedBlendEntry(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x20489F9AB36CC0E3, 0x20499874219E35EE), new AdvancedBlendEntry(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xBB176935E5EE05BF, 0x95B26D4D30EA7A14), new AdvancedBlendEntry(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x5FF9393C908ACFED, 0x068B0BD875773ABF), new AdvancedBlendEntry(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.2605f, 0.2605f, 0.2605f), new RgbFloat(-0.7817f, -0.7817f, -0.7817f), new RgbFloat(0.3022f, 0.3022f, 0.3022f), new RgbFloat(0.2192f, 0.2192f, 0.2192f), new RgbFloat(0.25f, 0.25f, 0.25f), new RgbFloat(16f, 16f, 16f), new RgbFloat(12f, 12f, 12f), new RgbFloat(3f, 3f, 3f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x03181F8711C9802C, 0x6B02C7C6B224FE7B), new AdvancedBlendEntry(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x2EE2209021F6B977, 0xF3AFA1491B8B89FC), new AdvancedBlendEntry(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xD8BA4DD2EDE4DC9E, 0x01006114977CF715), new AdvancedBlendEntry(AdvancedBlendOp.Invert, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0xD156B99835A2D8ED, 0x2D0BEE9E135EA7A7), new AdvancedBlendEntry(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x20CE8C898ED4BE27, 0x1514900B6F5E8F66), new AdvancedBlendEntry(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xCDE5F743820BA2D9, 0x917845FE2ECB083D), new AdvancedBlendEntry(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xEB03DF4A0C1D14CD, 0xBAE2E831C6E8FFE4), new AdvancedBlendEntry(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x1DC9E49AABC779AC, 0x4053A1441EB713D3), new AdvancedBlendEntry(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(2f, 2f, 2f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xFBDEF776248F7B3E, 0xE05EEFD65AC47CB7), new AdvancedBlendEntry(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x415A1A48E03AA6E7, 0x046D7EE33CA46B9A), new AdvancedBlendEntry(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Disjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x59A6901EC9BB2041, 0x2F3E19CE5EEC3EBE), new AdvancedBlendEntry(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x044B2B6E105221DA, 0x3089BBC033F994AF), new AdvancedBlendEntry(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x374A5A24AA8E6CC5, 0x29930FAA6215FA2B), new AdvancedBlendEntry(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x30CD0F7AF0CF26F9, 0x06CCA6744DE7DCF5), new AdvancedBlendEntry(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Disjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x1A6C9A1F6FE494A5, 0xA0CFAF77617E54DD), new AdvancedBlendEntry(AdvancedBlendOp.Src, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+            { new Hash128(0x081AF6DAAB1C8717, 0xBFEDCE59AE3DC9AC), new AdvancedBlendEntry(AdvancedBlendOp.Dst, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x3518E44573AB68BA, 0xC96EE71AF9F8F546), new AdvancedBlendEntry(AdvancedBlendOp.SrcOver, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0xF89E81FE8D73C96F, 0x4583A04577A0F21C), new AdvancedBlendEntry(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0xDF4026421CB61119, 0x14115A1F5139AFC7), new AdvancedBlendEntry(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MinimumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x91A20262C3E3A695, 0x0B3A102BFCDC6B1C), new AdvancedBlendEntry(AdvancedBlendOp.DstIn, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MinimumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x44F4C7CCFEB9EBFA, 0xF68394E6D56E5C2F), new AdvancedBlendEntry(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xB89F17C7021E9760, 0x430357EE0F7188EF), new AdvancedBlendEntry(AdvancedBlendOp.DstOut, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xDA2D20EA4242B8A0, 0x0D1EC05B72E3838F), new AdvancedBlendEntry(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x855DFEE1208D11B9, 0x77C6E3DDCFE30B85), new AdvancedBlendEntry(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+            { new Hash128(0x9B3808439683FD58, 0x123DCBE4705AB25E), new AdvancedBlendEntry(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xA42CF045C248A00A, 0x0C6C63C24EA0B0C1), new AdvancedBlendEntry(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x320A83B6D00C8059, 0x796EDAB3EB7314BC), new AdvancedBlendEntry(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x45253AC9ABFFC613, 0x8F92EA70195FB573), new AdvancedBlendEntry(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x1A5D263B588274B6, 0x167D305F6C794179), new AdvancedBlendEntry(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x709C1A837FE966AC, 0x75D8CE49E8A78EDB), new AdvancedBlendEntry(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x8265C26F85E4145F, 0x932E6CCBF37CB600), new AdvancedBlendEntry(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x3F252B3FEF983F27, 0x9370D7EEFEFA1A9E), new AdvancedBlendEntry(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x66A334A4AEA41078, 0xCB52254E1E395231), new AdvancedBlendEntry(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0xFDD05C53B25F0035, 0xB7E3ECEE166C222F), new AdvancedBlendEntry(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.2605f, 0.2605f, 0.2605f), new RgbFloat(-0.7817f, -0.7817f, -0.7817f), new RgbFloat(0.3022f, 0.3022f, 0.3022f), new RgbFloat(0.2192f, 0.2192f, 0.2192f), new RgbFloat(0.25f, 0.25f, 0.25f), new RgbFloat(16f, 16f, 16f), new RgbFloat(12f, 12f, 12f), new RgbFloat(3f, 3f, 3f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x25D932A77FFED81A, 0xA50D797B0FCA94E8), new AdvancedBlendEntry(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x4A953B6F5F7D341C, 0xDC05CFB50DDB5DC1), new AdvancedBlendEntry(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x838CB660C4F41F6D, 0x9E7D958697543495), new AdvancedBlendEntry(AdvancedBlendOp.Invert, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x4DF6EC1348A8F797, 0xA128E0CD69DB5A64), new AdvancedBlendEntry(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x178CDFAB9A015295, 0x2BF40EA72E596D57), new AdvancedBlendEntry(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x338FC99050E56AFD, 0x2AF41CF82BE602BF), new AdvancedBlendEntry(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x62E02ED60D1E978E, 0xBF726B3E68C11E4D), new AdvancedBlendEntry(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0xFBAF92DD4C101502, 0x7AF2EDA6596B819D), new AdvancedBlendEntry(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(2f, 2f, 2f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x0EF1241F65D4B50A, 0xE8D85DFA6AEDDB84), new AdvancedBlendEntry(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x77FE024B5C9D4A18, 0xF19D48A932F6860F), new AdvancedBlendEntry(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Conjoint, true, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x9C88CBFA2E09D857, 0x0A0361704CBEEE1D), new AdvancedBlendEntry(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x5B94127FA190E640, 0x8D1FEFF837A91268), new AdvancedBlendEntry(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0xB9C9105B7E063DDB, 0xF6A70E1D511B96FD), new AdvancedBlendEntry(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0xF0751AAE332B3ED1, 0xC40146F5C83C2533), new AdvancedBlendEntry(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Conjoint, true, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x579EB12F595F75AD, 0x151BF0504703B81B), new AdvancedBlendEntry(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0xF9CA152C03AC8C62, 0x1581336205E5CF47), new AdvancedBlendEntry(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.DstAlphaGl, BlendFactor.ZeroGl)) },
+            { new Hash128(0x98ACD8BB5E195D0F, 0x91F937672BE899F0), new AdvancedBlendEntry(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneMinusDstAlphaGl, BlendFactor.ZeroGl)) },
+            { new Hash128(0xBF97F10FC301F44C, 0x75721789F0D48548), new AdvancedBlendEntry(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x1B982263B8B08A10, 0x3350C76E2E1B27DF), new AdvancedBlendEntry(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+            { new Hash128(0xFF20AC79F64EDED8, 0xAF9025B2D97B9273), new AdvancedBlendEntry(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneMinusDstAlphaGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x9FFD986600FB112F, 0x384FDDF4E060139A), new AdvancedBlendEntry(AdvancedBlendOp.PlusClamped, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x0425E40B5B8B3B52, 0x5880CBED7CAB631C), new AdvancedBlendEntry(AdvancedBlendOp.PlusClampedAlpha, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x16DAC8593F28623A, 0x233DBC82325B8AED), new AdvancedBlendEntry(AdvancedBlendOp.PlusDarker, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xB37E5F234B9F0948, 0xD5F957A2ECD98FD6), new AdvancedBlendEntry(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0xCA0FDADD1D20DBE3, 0x1A5C15CCBF1AC538), new AdvancedBlendEntry(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x1C48304D73A9DF3A, 0x891DB93FA36E3450), new AdvancedBlendEntry(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x53200F2279B7FA39, 0x051C2462EBF6789C), new AdvancedBlendEntry(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0xB88BFB80714DCD5C, 0xEBD6938D744E6A41), new AdvancedBlendEntry(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0xE33DC2A25FC1A976, 0x08B3DBB1F3027D45), new AdvancedBlendEntry(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0xCE97E71615370316, 0xE131AE49D3A4D62B), new AdvancedBlendEntry(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0xE059FD265149B256, 0x94AF817AC348F61F), new AdvancedBlendEntry(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x16D31333D477E231, 0x9A98AAC84F72CC62), new AdvancedBlendEntry(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.2605f, 0.2605f, 0.2605f), new RgbFloat(-0.7817f, -0.7817f, -0.7817f), new RgbFloat(0.3022f, 0.3022f, 0.3022f), new RgbFloat(0.2192f, 0.2192f, 0.2192f), new RgbFloat(0.25f, 0.25f, 0.25f), new RgbFloat(16f, 16f, 16f), new RgbFloat(12f, 12f, 12f), new RgbFloat(3f, 3f, 3f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x47FC3B0776366D3C, 0xE96D9BD83B277874), new AdvancedBlendEntry(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x7230401E3FEA1F3B, 0xF0D15F05D3D1E309), new AdvancedBlendEntry(AdvancedBlendOp.Minus, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.ReverseSubtractGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x188212F9303742F5, 0x100C51CB96E03591), new AdvancedBlendEntry(AdvancedBlendOp.MinusClamped, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x52B755D296B44DC5, 0x4003B87275625973), new AdvancedBlendEntry(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0xD873ED973ADF7EAD, 0x73E68B57D92034E7), new AdvancedBlendEntry(AdvancedBlendOp.Contrast, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(2f, 2f, 2f), new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x471F9FA34B945ACB, 0x10524D1410B3C402), new AdvancedBlendEntry(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x99F569454EA0EF32, 0x6FC70A8B3A07DC8B), new AdvancedBlendEntry(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x5AD55F950067AC7E, 0x4BA60A4FBABDD0AC), new AdvancedBlendEntry(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x03FF2C858C9C4C5B, 0xE95AE7F561FB60E9), new AdvancedBlendEntry(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x6DC0E510C7BCF9D2, 0xAE805D7CECDCB5C1), new AdvancedBlendEntry(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(2f, 2f, 2f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x44832332CED5C054, 0x2F8D5536C085B30A), new AdvancedBlendEntry(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x4AB4D387618AC51F, 0x495B46E0555F4B32), new AdvancedBlendEntry(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x99282B49405A01A8, 0xD6FA93F864F24A8E), new AdvancedBlendEntry(AdvancedBlendOp.Red, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x37B30C1064FBD23E, 0x5D068366F42317C2), new AdvancedBlendEntry(AdvancedBlendOp.Green, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x760FAE9D59E04BC2, 0xA40AD483EA01435E), new AdvancedBlendEntry(AdvancedBlendOp.Blue, AdvancedBlendOverlap.Uncorrelated, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0xE786950FD9D1C6EF, 0xF9FDD5AF6451D239), new AdvancedBlendEntry(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x052458BB4788B0CA, 0x8AC58FDCA1F45EF5), new AdvancedBlendEntry(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x6AFC3837D1D31920, 0xB9D49C2FE49642C6), new AdvancedBlendEntry(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0xAFC2911949317E01, 0xD5B63636F5CB3422), new AdvancedBlendEntry(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Uncorrelated, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneMinusSrcAlphaGl)) },
+            { new Hash128(0x13B46DF507CC2C53, 0x86DE26517E6BF0A7), new AdvancedBlendEntry(AdvancedBlendOp.Src, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+            { new Hash128(0x5C372442474BE410, 0x79ECD3C0C496EF2E), new AdvancedBlendEntry(AdvancedBlendOp.SrcOver, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x74AAB45DBF5336E9, 0x01BFC4E181DAD442), new AdvancedBlendEntry(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x43239E282A36C85C, 0x36FB65560E46AD0F), new AdvancedBlendEntry(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x1A3BA8A7583B8F7A, 0xE64E41D548033180), new AdvancedBlendEntry(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x32BBB9859E9B565D, 0x3D5CE94FE55F18B5), new AdvancedBlendEntry(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0xD947A0766AE3C0FC, 0x391E5D53E86F4ED6), new AdvancedBlendEntry(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+            { new Hash128(0xBD9A7C08BDFD8CE6, 0x905407634901355E), new AdvancedBlendEntry(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x8395475BCB0D7A8C, 0x48AF5DD501D44A70), new AdvancedBlendEntry(AdvancedBlendOp.Plus, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x80AAC23FEBD4A3E5, 0xEA8C70F0B4DE52DE), new AdvancedBlendEntry(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x2F3AD1B0F1B3FD09, 0xC0EBC784BFAB8EA3), new AdvancedBlendEntry(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x52B54032F2F70BFF, 0xC941D6FDED674765), new AdvancedBlendEntry(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xCA7B86F72EC6A99B, 0x55868A131AFE359E), new AdvancedBlendEntry(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x377919B60BD133CA, 0x0FD611627664EF40), new AdvancedBlendEntry(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x9D4A0C5EE1153887, 0x7B869EBA218C589B), new AdvancedBlendEntry(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x311F2A858545D123, 0xB4D09C802480AD62), new AdvancedBlendEntry(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xCF78AA6A83AFA689, 0x9DC48B0C2182A3E1), new AdvancedBlendEntry(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xC3018CD6F1CF62D1, 0x016E32DD9087B1BB), new AdvancedBlendEntry(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.2605f, 0.2605f, 0.2605f), new RgbFloat(-0.7817f, -0.7817f, -0.7817f), new RgbFloat(0.3022f, 0.3022f, 0.3022f), new RgbFloat(0.2192f, 0.2192f, 0.2192f), new RgbFloat(0.25f, 0.25f, 0.25f), new RgbFloat(16f, 16f, 16f), new RgbFloat(12f, 12f, 12f), new RgbFloat(3f, 3f, 3f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x9CB62CE0E956EE29, 0x0FB67F503E60B3AD), new AdvancedBlendEntry(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x3589A13C16EF3BFA, 0x15B29BFC91F3BDFB), new AdvancedBlendEntry(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x3502CA5FB7529917, 0xFA51BFD0D1688071), new AdvancedBlendEntry(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x62ADC25AD6D0A923, 0x76CB6D238276D3A3), new AdvancedBlendEntry(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x09FDEB1116A9D52C, 0x85BB8627CD5C2733), new AdvancedBlendEntry(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x0709FED1B65E18EB, 0x5BC3AA4D99EC19CF), new AdvancedBlendEntry(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xB18D28AE5DE4C723, 0xE820AA2B75C9C02E), new AdvancedBlendEntry(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(2f, 2f, 2f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x6743C51621497480, 0x4B164E40858834AE), new AdvancedBlendEntry(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x63D1E181E34A2944, 0x1AE292C9D9F12819), new AdvancedBlendEntry(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Disjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x079523298250BFF6, 0xC0C793510603CDB5), new AdvancedBlendEntry(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x4C9D0A973C805EA6, 0xD1FF59AD5156B93C), new AdvancedBlendEntry(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x1E914678F3057BCD, 0xD503AE389C12D229), new AdvancedBlendEntry(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0x9FDBADE5556C5311, 0x03F0CBC798FC5C94), new AdvancedBlendEntry(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Disjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xE39451534635403C, 0x606CC1CA1F452388), new AdvancedBlendEntry(AdvancedBlendOp.Src, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+            { new Hash128(0x1D39F0F0A1008AA6, 0xBFDF2B97E6C3F125), new AdvancedBlendEntry(AdvancedBlendOp.SrcOver, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0xDB81BED30D5BDBEA, 0xAF0B2856EB93AD2C), new AdvancedBlendEntry(AdvancedBlendOp.DstOver, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x83F69CCF1D0A79B6, 0x70D31332797430AC), new AdvancedBlendEntry(AdvancedBlendOp.SrcIn, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MinimumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x7B87F807AB7A8F5C, 0x1241A2A01FB31771), new AdvancedBlendEntry(AdvancedBlendOp.SrcOut, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xF557172E20D5272D, 0xC1961F8C7A5D2820), new AdvancedBlendEntry(AdvancedBlendOp.SrcAtop, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0xA8476B3944DBBC9B, 0x84A2F6AF97B15FDF), new AdvancedBlendEntry(AdvancedBlendOp.DstAtop, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.OneGl, BlendFactor.ZeroGl)) },
+            { new Hash128(0x3259602B55414DA3, 0x72AACCC00B5A9D10), new AdvancedBlendEntry(AdvancedBlendOp.Xor, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, 0, 0, 0)) },
+            { new Hash128(0xC0CB8C10F36EDCD6, 0x8C2D088AD8191E1C), new AdvancedBlendEntry(AdvancedBlendOp.Multiply, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x81806C451C6255EF, 0x5AA8AC9A08941A15), new AdvancedBlendEntry(AdvancedBlendOp.Screen, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0xE55A6537F4568198, 0xCA8735390B799B19), new AdvancedBlendEntry(AdvancedBlendOp.Overlay, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x5C044BA14536DDA3, 0xBCE0123ED7D510EC), new AdvancedBlendEntry(AdvancedBlendOp.Darken, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x6788346C405BE130, 0x372A4BB199C01F9F), new AdvancedBlendEntry(AdvancedBlendOp.Lighten, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x510EDC2A34E2856B, 0xE1727A407E294254), new AdvancedBlendEntry(AdvancedBlendOp.ColorDodge, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x4B7BE01BD398C7A8, 0x5BFF79BC00672C18), new AdvancedBlendEntry(AdvancedBlendOp.ColorBurn, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x213B43845540CFEC, 0xDA857411CF1CCFCE), new AdvancedBlendEntry(AdvancedBlendOp.HardLight, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x765AFA6732E783F1, 0x8F1CABF1BC78A014), new AdvancedBlendEntry(AdvancedBlendOp.SoftLight, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.2605f, 0.2605f, 0.2605f), new RgbFloat(-0.7817f, -0.7817f, -0.7817f), new RgbFloat(0.3022f, 0.3022f, 0.3022f), new RgbFloat(0.2192f, 0.2192f, 0.2192f), new RgbFloat(0.25f, 0.25f, 0.25f), new RgbFloat(16f, 16f, 16f), new RgbFloat(12f, 12f, 12f), new RgbFloat(3f, 3f, 3f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0xA4A5DE1CC06F6CB1, 0xA0634A0011001709), new AdvancedBlendEntry(AdvancedBlendOp.Difference, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x81F32BD8816EA796, 0x697EE86683165170), new AdvancedBlendEntry(AdvancedBlendOp.Exclusion, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0xB870C209EAA5F092, 0xAF5FD923909CAA1F), new AdvancedBlendEntry(AdvancedBlendOp.InvertRGB, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.AddGl, BlendFactor.ZeroGl, BlendFactor.OneGl)) },
+            { new Hash128(0x3649A9F5C936FB83, 0xDD7C834897AA182A), new AdvancedBlendEntry(AdvancedBlendOp.LinearDodge, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0xD72A2B1097A5995C, 0x3D41B2763A913654), new AdvancedBlendEntry(AdvancedBlendOp.LinearBurn, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x551E212B9F6C454A, 0xB0DFA05BEB3C37FA), new AdvancedBlendEntry(AdvancedBlendOp.VividLight, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.5f, 0.5f, 0.5f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x681B5A313B7416BF, 0xCB1CBAEEB4D81500), new AdvancedBlendEntry(AdvancedBlendOp.LinearLight, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(2f, 2f, 2f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x9343A18BD4B16777, 0xEDB4AC1C8972C3A4), new AdvancedBlendEntry(AdvancedBlendOp.PinLight, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0xC960BF6D8519DE28, 0x78D8557FD405D119), new AdvancedBlendEntry(AdvancedBlendOp.HardMix, AdvancedBlendOverlap.Conjoint, false, Array.Empty<RgbFloat>(), new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x65A7B01FDC73A46C, 0x297E096ED5CC4D8A), new AdvancedBlendEntry(AdvancedBlendOp.HslHue, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0xD9C99BA4A6CDC13B, 0x3CFF0ACEDC2EE150), new AdvancedBlendEntry(AdvancedBlendOp.HslSaturation, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x6BC00DA6EB922BD1, 0x5FD4C11F2A685234), new AdvancedBlendEntry(AdvancedBlendOp.HslColor, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+            { new Hash128(0x8652300E32D93050, 0x9460E7B449132371), new AdvancedBlendEntry(AdvancedBlendOp.HslLuminosity, AdvancedBlendOverlap.Conjoint, false, new[] { new RgbFloat(0.3f, 0.59f, 0.11f) }, new FixedFunctionAlpha(BlendUcodeEnable.EnableRGB, BlendOp.MaximumGl, BlendFactor.OneGl, BlendFactor.OneGl)) },
+        };
+    }
+}
\ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendUcode.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendUcode.cs
new file mode 100644
index 00000000..f06b4bf7
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/AdvancedBlendUcode.cs
@@ -0,0 +1,126 @@
+using Ryujinx.Graphics.GAL;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed.Blender
+{
+    /// <summary>
+    /// Fixed function alpha state used for a advanced blend function.
+    /// </summary>
+    struct FixedFunctionAlpha
+    {
+        /// <summary>
+        /// Fixed function alpha state with alpha blending disabled.
+        /// </summary>
+        public static FixedFunctionAlpha Disabled => new FixedFunctionAlpha(BlendUcodeEnable.EnableRGBA, default, default, default);
+
+        /// <summary>
+        /// Individual enable bits for the RGB and alpha components.
+        /// </summary>
+        public BlendUcodeEnable Enable { get; }
+
+        /// <summary>
+        /// Alpha blend operation.
+        /// </summary>
+        public BlendOp AlphaOp { get; }
+
+        /// <summary>
+        /// Value multiplied with the blend source operand.
+        /// </summary>
+        public BlendFactor AlphaSrcFactor { get; }
+
+        /// <summary>
+        /// Value multiplied with the blend destination operand.
+        /// </summary>
+        public BlendFactor AlphaDstFactor { get; }
+
+        /// <summary>
+        /// Creates a new blend fixed function alpha state.
+        /// </summary>
+        /// <param name="enable">Individual enable bits for the RGB and alpha components</param>
+        /// <param name="alphaOp">Alpha blend operation</param>
+        /// <param name="alphaSrc">Value multiplied with the blend source operand</param>
+        /// <param name="alphaDst">Value multiplied with the blend destination operand</param>
+        public FixedFunctionAlpha(BlendUcodeEnable enable, BlendOp alphaOp, BlendFactor alphaSrc, BlendFactor alphaDst)
+        {
+            Enable = enable;
+            AlphaOp = alphaOp;
+            AlphaSrcFactor = alphaSrc;
+            AlphaDstFactor = alphaDst;
+        }
+
+        /// <summary>
+        /// Creates a new blend fixed function alpha state.
+        /// </summary>
+        /// <param name="alphaOp">Alpha blend operation</param>
+        /// <param name="alphaSrc">Value multiplied with the blend source operand</param>
+        /// <param name="alphaDst">Value multiplied with the blend destination operand</param>
+        public FixedFunctionAlpha(BlendOp alphaOp, BlendFactor alphaSrc, BlendFactor alphaDst) : this(BlendUcodeEnable.EnableRGB, alphaOp, alphaSrc, alphaDst)
+        {
+        }
+    }
+
+    /// <summary>
+    /// Blend microcode assembly function delegate.
+    /// </summary>
+    /// <param name="asm">Assembler</param>
+    /// <returns>Fixed function alpha state for the microcode</returns>
+    delegate FixedFunctionAlpha GenUcodeFunc(ref UcodeAssembler asm);
+
+    /// <summary>
+    /// Advanced blend microcode state.
+    /// </summary>
+    struct AdvancedBlendUcode
+    {
+        /// <summary>
+        /// Advanced blend operation.
+        /// </summary>
+        public AdvancedBlendOp Op { get; }
+
+        /// <summary>
+        /// Advanced blend overlap mode.
+        /// </summary>
+        public AdvancedBlendOverlap Overlap { get; }
+
+        /// <summary>
+        /// Whenever the source input is pre-multiplied.
+        /// </summary>
+        public bool SrcPreMultiplied { get; }
+
+        /// <summary>
+        /// Fixed function alpha state.
+        /// </summary>
+        public FixedFunctionAlpha Alpha { get; }
+
+        /// <summary>
+        /// Microcode.
+        /// </summary>
+        public uint[] Code { get; }
+
+        /// <summary>
+        /// Constants used by the microcode.
+        /// </summary>
+        public RgbFloat[] Constants { get; }
+
+        /// <summary>
+        /// Creates a new advanced blend state.
+        /// </summary>
+        /// <param name="op">Advanced blend operation</param>
+        /// <param name="overlap">Advanced blend overlap mode</param>
+        /// <param name="srcPreMultiplied">Whenever the source input is pre-multiplied</param>
+        /// <param name="genFunc">Function that will generate the advanced blend microcode</param>
+        public AdvancedBlendUcode(
+            AdvancedBlendOp op,
+            AdvancedBlendOverlap overlap,
+            bool srcPreMultiplied,
+            GenUcodeFunc genFunc)
+        {
+            Op = op;
+            Overlap = overlap;
+            SrcPreMultiplied = srcPreMultiplied;
+
+            UcodeAssembler asm = new UcodeAssembler();
+            Alpha = genFunc(ref asm);
+            Code = asm.GetCode();
+            Constants = asm.GetConstants();
+        }
+    }
+}
\ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/UcodeAssembler.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/UcodeAssembler.cs
new file mode 100644
index 00000000..f854787e
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/Blender/UcodeAssembler.cs
@@ -0,0 +1,305 @@
+using System;
+using System.Collections.Generic;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed.Blender
+{
+    /// <summary>
+    /// Blend microcode instruction.
+    /// </summary>
+    enum Instruction
+    {
+        Mmadd = 0,
+        Mmsub = 1,
+        Min = 2,
+        Max = 3,
+        Rcp = 4,
+        Add = 5,
+        Sub = 6
+    }
+
+    /// <summary>
+    /// Blend microcode condition code.
+    /// </summary>
+    enum CC
+    {
+        F = 0,
+        T = 1,
+        EQ = 2,
+        NE = 3,
+        LT = 4,
+        LE = 5,
+        GT = 6,
+        GE = 7
+    }
+
+    /// <summary>
+    /// Blend microcode opend B or D value.
+    /// </summary>
+    enum OpBD
+    {
+        ConstantZero = 0x0,
+        ConstantOne = 0x1,
+        SrcRGB = 0x2,
+        SrcAAA = 0x3,
+        OneMinusSrcAAA = 0x4,
+        DstRGB = 0x5,
+        DstAAA = 0x6,
+        OneMinusDstAAA = 0x7,
+        Temp0 = 0x9,
+        Temp1 = 0xa,
+        Temp2 = 0xb,
+        PBR = 0xc,
+        ConstantRGB = 0xd
+    }
+
+    /// <summary>
+    /// Blend microcode operand A or C value.
+    /// </summary>
+    enum OpAC
+    {
+        SrcRGB = 0,
+        DstRGB = 1,
+        SrcAAA = 2,
+        DstAAA = 3,
+        Temp0 = 4,
+        Temp1 = 5,
+        Temp2 = 6,
+        PBR = 7
+    }
+
+    /// <summary>
+    /// Blend microcode destination operand.
+    /// </summary>
+    enum OpDst
+    {
+        Temp0 = 0,
+        Temp1 = 1,
+        Temp2 = 2,
+        PBR = 3
+    }
+
+    /// <summary>
+    /// Blend microcode input swizzle.
+    /// </summary>
+    enum Swizzle
+    {
+        RGB = 0,
+        GBR = 1,
+        RRR = 2,
+        GGG = 3,
+        BBB = 4,
+        RToA = 5
+    }
+
+    /// <summary>
+    /// Blend microcode output components.
+    /// </summary>
+    enum WriteMask
+    {
+        RGB = 0,
+        R = 1,
+        G = 2,
+        B = 3
+    }
+
+    /// <summary>
+    /// Floating-point RGB color values.
+    /// </summary>
+    struct RgbFloat
+    {
+        /// <summary>
+        /// Red component value.
+        /// </summary>
+        public float R { get; }
+
+        /// <summary>
+        /// Green component value.
+        /// </summary>
+        public float G { get; }
+
+        /// <summary>
+        /// Blue component value.
+        /// </summary>
+        public float B { get; }
+
+        /// <summary>
+        /// Creates a new floating-point RGB value.
+        /// </summary>
+        /// <param name="r">Red component value</param>
+        /// <param name="g">Green component value</param>
+        /// <param name="b">Blue component value</param>
+        public RgbFloat(float r, float g, float b)
+        {
+            R = r;
+            G = g;
+            B = b;
+        }
+    }
+
+    /// <summary>
+    /// Blend microcode destination operand, including swizzle, write mask and condition code update flag.
+    /// </summary>
+    struct Dest
+    {
+        public static Dest Temp0 => new Dest(OpDst.Temp0, Swizzle.RGB, WriteMask.RGB, false);
+        public static Dest Temp1 => new Dest(OpDst.Temp1, Swizzle.RGB, WriteMask.RGB, false);
+        public static Dest Temp2 => new Dest(OpDst.Temp2, Swizzle.RGB, WriteMask.RGB, false);
+        public static Dest PBR => new Dest(OpDst.PBR, Swizzle.RGB, WriteMask.RGB, false);
+
+        public Dest GBR => new Dest(Dst, Swizzle.GBR, WriteMask, WriteCC);
+        public Dest RRR => new Dest(Dst, Swizzle.RRR, WriteMask, WriteCC);
+        public Dest GGG => new Dest(Dst, Swizzle.GGG, WriteMask, WriteCC);
+        public Dest BBB => new Dest(Dst, Swizzle.BBB, WriteMask, WriteCC);
+        public Dest RToA => new Dest(Dst, Swizzle.RToA, WriteMask, WriteCC);
+
+        public Dest R => new Dest(Dst, Swizzle, WriteMask.R, WriteCC);
+        public Dest G => new Dest(Dst, Swizzle, WriteMask.G, WriteCC);
+        public Dest B => new Dest(Dst, Swizzle, WriteMask.B, WriteCC);
+
+        public Dest CC => new Dest(Dst, Swizzle, WriteMask, true);
+
+        public OpDst Dst { get; }
+        public Swizzle Swizzle { get; }
+        public WriteMask WriteMask { get; }
+        public bool WriteCC { get; }
+
+        /// <summary>
+        /// Creates a new blend microcode destination operand.
+        /// </summary>
+        /// <param name="dst">Operand</param>
+        /// <param name="swizzle">Swizzle</param>
+        /// <param name="writeMask">Write maks</param>
+        /// <param name="writeCC">Indicates if condition codes should be updated</param>
+        public Dest(OpDst dst, Swizzle swizzle, WriteMask writeMask, bool writeCC)
+        {
+            Dst = dst;
+            Swizzle = swizzle;
+            WriteMask = writeMask;
+            WriteCC = writeCC;
+        }
+    }
+
+    /// <summary>
+    /// Blend microcode operaiton.
+    /// </summary>
+    struct UcodeOp
+    {
+        public readonly uint Word;
+
+        /// <summary>
+        /// Creates a new blend microcode operation.
+        /// </summary>
+        /// <param name="cc">Condition code that controls whenever the operation is executed or not</param>
+        /// <param name="inst">Instruction</param>
+        /// <param name="constIndex">Index on the constant table of the constant used by any constant operand</param>
+        /// <param name="dest">Destination operand</param>
+        /// <param name="srcA">First input operand</param>
+        /// <param name="srcB">Second input operand</param>
+        /// <param name="srcC">Third input operand</param>
+        /// <param name="srcD">Fourth input operand</param>
+        public UcodeOp(CC cc, Instruction inst, int constIndex, Dest dest, OpAC srcA, OpBD srcB, OpAC srcC, OpBD srcD)
+        {
+            Word = (uint)cc |
+                ((uint)inst << 3) |
+                ((uint)constIndex << 6) |
+                ((uint)srcA << 9) |
+                ((uint)srcB << 12) |
+                ((uint)srcC << 16) |
+                ((uint)srcD << 19) |
+                ((uint)dest.Swizzle << 23) |
+                ((uint)dest.WriteMask << 26) |
+                ((uint)dest.Dst << 28) |
+                (dest.WriteCC ? (1u << 31) : 0);
+        }
+    }
+
+    /// <summary>
+    /// Blend microcode assembler.
+    /// </summary>
+    struct UcodeAssembler
+    {
+        private List<uint> _code;
+        private RgbFloat[] _constants;
+        private int _constantIndex;
+
+        public void Mul(CC cc, Dest dest, OpAC srcA, OpBD srcB)
+        {
+            Assemble(cc, Instruction.Mmadd, dest, srcA, srcB, OpAC.SrcRGB, OpBD.ConstantZero);
+        }
+
+        public void Madd(CC cc, Dest dest, OpAC srcA, OpBD srcB, OpAC srcC)
+        {
+            Assemble(cc, Instruction.Mmadd, dest, srcA, srcB, srcC, OpBD.ConstantOne);
+        }
+
+        public void Mmadd(CC cc, Dest dest, OpAC srcA, OpBD srcB, OpAC srcC, OpBD srcD)
+        {
+            Assemble(cc, Instruction.Mmadd, dest, srcA, srcB, srcC, srcD);
+        }
+
+        public void Mmsub(CC cc, Dest dest, OpAC srcA, OpBD srcB, OpAC srcC, OpBD srcD)
+        {
+            Assemble(cc, Instruction.Mmsub, dest, srcA, srcB, srcC, srcD);
+        }
+
+        public void Min(CC cc, Dest dest, OpAC srcA, OpBD srcB)
+        {
+            Assemble(cc, Instruction.Min, dest, srcA, srcB, OpAC.SrcRGB, OpBD.ConstantZero);
+        }
+
+        public void Max(CC cc, Dest dest, OpAC srcA, OpBD srcB)
+        {
+            Assemble(cc, Instruction.Max, dest, srcA, srcB, OpAC.SrcRGB, OpBD.ConstantZero);
+        }
+
+        public void Rcp(CC cc, Dest dest, OpAC srcA)
+        {
+            Assemble(cc, Instruction.Rcp, dest, srcA, OpBD.ConstantZero, OpAC.SrcRGB, OpBD.ConstantZero);
+        }
+
+        public void Mov(CC cc, Dest dest, OpBD srcB)
+        {
+            Assemble(cc, Instruction.Add, dest, OpAC.SrcRGB, srcB, OpAC.SrcRGB, OpBD.ConstantZero);
+        }
+
+        public void Add(CC cc, Dest dest, OpBD srcB, OpBD srcD)
+        {
+            Assemble(cc, Instruction.Add, dest, OpAC.SrcRGB, srcB, OpAC.SrcRGB, srcD);
+        }
+
+        public void Sub(CC cc, Dest dest, OpBD srcB, OpBD srcD)
+        {
+            Assemble(cc, Instruction.Sub, dest, OpAC.SrcRGB, srcB, OpAC.SrcRGB, srcD);
+        }
+
+        private void Assemble(CC cc, Instruction inst, Dest dest, OpAC srcA, OpBD srcB, OpAC srcC, OpBD srcD)
+        {
+            (_code ??= new List<uint>()).Add(new UcodeOp(cc, inst, _constantIndex, dest, srcA, srcB, srcC, srcD).Word);
+        }
+
+        public void SetConstant(int index, float r, float g, float b)
+        {
+            if (_constants == null)
+            {
+                _constants = new RgbFloat[index + 1];
+            }
+            else if (_constants.Length <= index)
+            {
+                Array.Resize(ref _constants, index + 1);
+            }
+
+            _constants[index] = new RgbFloat(r, g, b);
+            _constantIndex = index;
+        }
+
+        public uint[] GetCode()
+        {
+            return _code?.ToArray();
+        }
+
+        public RgbFloat[] GetConstants()
+        {
+            return _constants;
+        }
+    }
+}
\ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/ConditionalRendering.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ConditionalRendering.cs
new file mode 100644
index 00000000..a6b62a4a
--- /dev/null
+++ b/src/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, true) != 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, true);
+                ulong y = memoryManager.Read<ulong>(gpuVa + 16, true);
+
+                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);
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/ConstantBufferUpdater.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ConstantBufferUpdater.cs
new file mode 100644
index 00000000..5c936616
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ConstantBufferUpdater.cs
@@ -0,0 +1,183 @@
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Constant buffer updater.
+    /// </summary>
+    class ConstantBufferUpdater
+    {
+        private const int UniformDataCacheSize = 512;
+
+        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;
+        private int _ubIndex = 0;
+        private int[] _ubData = new int[UniformDataCacheSize];
+
+        /// <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;
+
+                Span<byte> data = MemoryMarshal.Cast<int, byte>(_ubData.AsSpan(0, (int)(_ubByteCount / 4)));
+
+                if (memoryManager.Physical.WriteWithRedundancyCheck(_ubBeginCpuAddress, data))
+                {
+                    memoryManager.Physical.BufferCache.ForceDirty(memoryManager, _ubFollowUpAddress - _ubByteCount, _ubByteCount);
+                }
+
+                _ubFollowUpAddress = 0;
+                _ubIndex = 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 || _ubIndex == _ubData.Length)
+            {
+                FlushUboDirty();
+
+                _ubByteCount = 0;
+                _ubBeginCpuAddress = _channel.MemoryManager.Translate(address);
+            }
+
+            _ubData[_ubIndex++] = argument;
+
+            _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 || _ubIndex + data.Length > _ubData.Length)
+            {
+                FlushUboDirty();
+
+                _ubByteCount = 0;
+                _ubBeginCpuAddress = _channel.MemoryManager.Translate(address);
+            }
+
+            data.CopyTo(_ubData.AsSpan(_ubIndex));
+            _ubIndex += data.Length;
+
+            _ubFollowUpAddress = address + size;
+            _ubByteCount += size;
+
+            _state.State.UniformBufferState.Offset += data.Length * 4;
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawManager.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawManager.cs
new file mode 100644
index 00000000..7438ba03
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawManager.cs
@@ -0,0 +1,856 @@
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.Types;
+using Ryujinx.Graphics.Gpu.Memory;
+using System;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Draw manager.
+    /// </summary>
+    class DrawManager
+    {
+        // Since we don't know the index buffer size for indirect draws,
+        // we must assume a minimum and maximum size and use that for buffer data update purposes.
+        private const int MinIndirectIndexCount = 0x10000;
+        private const int MaxIndirectIndexCount = 0x4000000;
+
+        private readonly GpuContext _context;
+        private readonly GpuChannel _channel;
+        private readonly DeviceStateWithShadow<ThreedClassState> _state;
+        private readonly DrawState _drawState;
+        private readonly SpecializationStateUpdater _currentSpecState;
+        private bool _topologySet;
+
+        private bool _instancedDrawPending;
+        private bool _instancedIndexed;
+        private bool _instancedIndexedInline;
+
+        private int _instancedFirstIndex;
+        private int _instancedFirstVertex;
+        private int _instancedFirstInstance;
+        private int _instancedIndexCount;
+        private int _instancedDrawStateFirst;
+        private int _instancedDrawStateCount;
+
+        private int _instanceIndex;
+
+        private const int VertexBufferFirstMethodOffset = 0x35d;
+        private const int IndexBufferCountMethodOffset = 0x5f8;
+
+        /// <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>
+        /// <param name="spec">Specialization state updater</param>
+        public DrawManager(GpuContext context, GpuChannel channel, DeviceStateWithShadow<ThreedClassState> state, DrawState drawState, SpecializationStateUpdater spec)
+        {
+            _context = context;
+            _channel = channel;
+            _state = state;
+            _drawState = drawState;
+            _currentSpecState = spec;
+        }
+
+        /// <summary>
+        /// Marks the entire state as dirty, forcing a full host state update before the next draw.
+        /// </summary>
+        public void ForceStateDirty()
+        {
+            _topologySet = false;
+        }
+
+        /// <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,
+                _state.State.VertexBufferDrawState.First,
+                _state.State.VertexBufferDrawState.Count);
+        }
+
+        /// <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>
+        /// <param name="drawFirstVertex">Index of the first vertex used on the draw</param>
+        /// <param name="drawVertexCount">Number of vertices used on the draw</param>
+        private void DrawEnd(ThreedClass engine, int firstIndex, int indexCount, int drawFirstVertex, int drawVertexCount)
+        {
+            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;
+            _drawState.DrawFirstVertex = drawFirstVertex;
+            _drawState.DrawVertexCount = drawVertexCount;
+            _currentSpecState.SetHasConstantBufferDrawParameters(false);
+
+            engine.UpdateState();
+
+            bool instanced = _drawState.VsUsesInstanceId || _drawState.IsAnyVbInstanced;
+
+            if (instanced)
+            {
+                _instancedDrawPending = true;
+
+                int ibCount = _drawState.IbStreamer.InlineIndexCount;
+
+                _instancedIndexed = _drawState.DrawIndexed;
+                _instancedIndexedInline = ibCount != 0;
+
+                _instancedFirstIndex = firstIndex;
+                _instancedFirstVertex = (int)_state.State.FirstVertex;
+                _instancedFirstInstance = (int)_state.State.FirstInstance;
+
+                _instancedIndexCount = ibCount != 0 ? ibCount : indexCount;
+
+                _instancedDrawStateFirst = drawFirstVertex;
+                _instancedDrawStateCount = drawVertexCount;
+
+                _drawState.DrawIndexed = false;
+
+                if (renderEnable == ConditionalRenderEnabled.Host)
+                {
+                    _context.Renderer.Pipeline.EndHostConditionalRendering();
+                }
+
+                return;
+            }
+
+            int firstInstance = (int)_state.State.FirstInstance;
+
+            int inlineIndexCount = _drawState.IbStreamer.GetAndResetInlineIndexCount(_context.Renderer);
+
+            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(drawVertexCount, 1, drawFirstVertex, 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;
+
+            PrimitiveType type = (PrimitiveType)(argument & 0xffff);
+            DrawBegin(incrementInstance, resetInstance, type);
+        }
+
+        /// <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="primitiveType">Primitive type</param>
+        private void DrawBegin(bool incrementInstance, bool resetInstance, PrimitiveType primitiveType)
+        {
+            if (incrementInstance)
+            {
+                _instanceIndex++;
+            }
+            else if (resetInstance)
+            {
+                PerformDeferredDraws();
+
+                _instanceIndex = 0;
+            }
+
+            PrimitiveTopology topology;
+
+            if (_state.State.PrimitiveTypeOverrideEnable)
+            {
+                PrimitiveTypeOverride typeOverride = _state.State.PrimitiveTypeOverride;
+                topology = typeOverride.Convert();
+            }
+            else
+            {
+                topology = primitiveType.Convert();
+            }
+
+            UpdateTopology(topology);
+        }
+
+        /// <summary>
+        /// Updates the current primitive topology if needed.
+        /// </summary>
+        /// <param name="topology">New primitive topology</param>
+        private void UpdateTopology(PrimitiveTopology topology)
+        {
+            if (_drawState.Topology != topology || !_topologySet)
+            {
+                _context.Renderer.Pipeline.SetPrimitiveTopology(topology);
+                _currentSpecState.SetTopology(topology);
+                _drawState.Topology = topology;
+                _topologySet = true;
+            }
+        }
+
+        /// <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;
+        }
+
+        // TODO: Verify if the index type is implied from the method that is called,
+        // or if it uses the state index type on hardware.
+
+        /// <summary>
+        /// Performs a indexed draw with 8-bit 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 DrawIndexBuffer8BeginEndInstanceFirst(ThreedClass engine, int argument)
+        {
+            DrawIndexBufferBeginEndInstance(engine, argument, false);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with 16-bit 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 DrawIndexBuffer16BeginEndInstanceFirst(ThreedClass engine, int argument)
+        {
+            DrawIndexBufferBeginEndInstance(engine, argument, false);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with 32-bit 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 DrawIndexBuffer32BeginEndInstanceFirst(ThreedClass engine, int argument)
+        {
+            DrawIndexBufferBeginEndInstance(engine, argument, false);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with 8-bit 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 DrawIndexBuffer8BeginEndInstanceSubsequent(ThreedClass engine, int argument)
+        {
+            DrawIndexBufferBeginEndInstance(engine, argument, true);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with 16-bit 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 DrawIndexBuffer16BeginEndInstanceSubsequent(ThreedClass engine, int argument)
+        {
+            DrawIndexBufferBeginEndInstance(engine, argument, true);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with 32-bit 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 DrawIndexBuffer32BeginEndInstanceSubsequent(ThreedClass engine, int argument)
+        {
+            DrawIndexBufferBeginEndInstance(engine, argument, true);
+        }
+
+        /// <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 DrawIndexBufferBeginEndInstance(ThreedClass engine, int argument, bool instanced)
+        {
+            DrawBegin(instanced, !instanced, (PrimitiveType)((argument >> 28) & 0xf));
+
+            int firstIndex = argument & 0xffff;
+            int indexCount = (argument >> 16) & 0xfff;
+
+            bool oldDrawIndexed = _drawState.DrawIndexed;
+
+            _drawState.DrawIndexed = true;
+            engine.ForceStateDirty(IndexBufferCountMethodOffset * 4);
+
+            DrawEnd(engine, firstIndex, indexCount, 0, 0);
+
+            _drawState.DrawIndexed = oldDrawIndexed;
+        }
+
+        /// <summary>
+        /// Performs a non-indexed draw with the specified topology, index and count.
+        /// </summary>
+        /// <param name="engine">3D engine where this method is being called</param>
+        /// <param name="argument">Method call argument</param>
+        public void DrawVertexArrayBeginEndInstanceFirst(ThreedClass engine, int argument)
+        {
+            DrawVertexArrayBeginEndInstance(engine, argument, false);
+        }
+
+        /// <summary>
+        /// Performs a non-indexed draw with the specified topology, index and count,
+        /// while incrementing the current instance.
+        /// </summary>
+        /// <param name="engine">3D engine where this method is being called</param>
+        /// <param name="argument">Method call argument</param>
+        public void DrawVertexArrayBeginEndInstanceSubsequent(ThreedClass engine, int argument)
+        {
+            DrawVertexArrayBeginEndInstance(engine, argument, true);
+        }
+
+        /// <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 DrawVertexArrayBeginEndInstance(ThreedClass engine, int argument, bool instanced)
+        {
+            DrawBegin(instanced, !instanced, (PrimitiveType)((argument >> 28) & 0xf));
+
+            int firstVertex = argument & 0xffff;
+            int vertexCount = (argument >> 16) & 0xfff;
+
+            bool oldDrawIndexed = _drawState.DrawIndexed;
+
+            _drawState.DrawIndexed = false;
+            engine.ForceStateDirty(VertexBufferFirstMethodOffset * 4);
+
+            DrawEnd(engine, 0, 0, firstVertex, vertexCount);
+
+            _drawState.DrawIndexed = oldDrawIndexed;
+        }
+
+        /// <summary>
+        /// Performs a texture draw with a source texture and sampler ID, along with source
+        /// and destination coordinates and sizes.
+        /// </summary>
+        /// <param name="engine">3D engine where this method is being called</param>
+        /// <param name="argument">Method call argument</param>
+        public void DrawTexture(ThreedClass engine, int argument)
+        {
+            static float FixedToFloat(int fixedValue)
+            {
+                return fixedValue * (1f / 4096);
+            }
+
+            float dstX0 = FixedToFloat(_state.State.DrawTextureDstX);
+            float dstY0 = FixedToFloat(_state.State.DrawTextureDstY);
+            float dstWidth = FixedToFloat(_state.State.DrawTextureDstWidth);
+            float dstHeight = FixedToFloat(_state.State.DrawTextureDstHeight);
+
+            // TODO: Confirm behaviour on hardware.
+            // When this is active, the origin appears to be on the bottom.
+            if (_state.State.YControl.HasFlag(YControl.NegateY))
+            {
+                dstY0 -= dstHeight;
+            }
+
+            float dstX1 = dstX0 + dstWidth;
+            float dstY1 = dstY0 + dstHeight;
+
+            float srcX0 = FixedToFloat(_state.State.DrawTextureSrcX);
+            float srcY0 = FixedToFloat(_state.State.DrawTextureSrcY);
+            float srcX1 = ((float)_state.State.DrawTextureDuDx / (1UL << 32)) * dstWidth + srcX0;
+            float srcY1 = ((float)_state.State.DrawTextureDvDy / (1UL << 32)) * dstHeight + srcY0;
+
+            engine.UpdateState(ulong.MaxValue & ~(1UL << StateUpdater.ShaderStateIndex));
+
+            _channel.TextureManager.UpdateRenderTargets();
+
+            int textureId = _state.State.DrawTextureTextureId;
+            int samplerId = _state.State.DrawTextureSamplerId;
+
+            (var texture, var sampler) = _channel.TextureManager.GetGraphicsTextureAndSampler(textureId, samplerId);
+
+            srcX0 *= texture.ScaleFactor;
+            srcY0 *= texture.ScaleFactor;
+            srcX1 *= texture.ScaleFactor;
+            srcY1 *= texture.ScaleFactor;
+
+            float dstScale = _channel.TextureManager.RenderTargetScale;
+
+            dstX0 *= dstScale;
+            dstY0 *= dstScale;
+            dstX1 *= dstScale;
+            dstY1 *= dstScale;
+
+            _context.Renderer.Pipeline.DrawTexture(
+                texture?.HostTexture,
+                sampler?.GetHostSampler(texture),
+                new Extents2DF(srcX0, srcY0, srcX1, srcY1),
+                new Extents2DF(dstX0, dstY0, dstX1, dstY1));
+        }
+
+        /// <summary>
+        /// Performs a indexed or non-indexed draw.
+        /// </summary>
+        /// <param name="engine">3D engine where this method is being called</param>
+        /// <param name="topology">Primitive topology</param>
+        /// <param name="count">Index count for indexed draws, vertex count for non-indexed draws</param>
+        /// <param name="instanceCount">Instance count</param>
+        /// <param name="firstIndex">First index on the index buffer for indexed draws, ignored for non-indexed draws</param>
+        /// <param name="firstVertex">First vertex on the vertex buffer</param>
+        /// <param name="firstInstance">First instance</param>
+        /// <param name="indexed">True if the draw is indexed, false otherwise</param>
+        public void Draw(
+            ThreedClass engine,
+            PrimitiveTopology topology,
+            int count,
+            int instanceCount,
+            int firstIndex,
+            int firstVertex,
+            int firstInstance,
+            bool indexed)
+        {
+            UpdateTopology(topology);
+
+            ConditionalRenderEnabled renderEnable = ConditionalRendering.GetRenderEnable(
+                _context,
+                _channel.MemoryManager,
+                _state.State.RenderEnableAddress,
+                _state.State.RenderEnableCondition);
+
+            if (renderEnable == ConditionalRenderEnabled.False)
+            {
+                _drawState.DrawIndexed = false;
+                return;
+            }
+
+            if (indexed)
+            {
+                _drawState.FirstIndex = firstIndex;
+                _drawState.IndexCount = count;
+                _state.State.FirstVertex = (uint)firstVertex;
+                engine.ForceStateDirty(IndexBufferCountMethodOffset * 4);
+            }
+            else
+            {
+                _drawState.DrawFirstVertex = firstVertex;
+                _drawState.DrawVertexCount = count;
+                engine.ForceStateDirty(VertexBufferFirstMethodOffset * 4);
+            }
+
+            _state.State.FirstInstance = (uint)firstInstance;
+
+            _drawState.DrawIndexed = indexed;
+            _currentSpecState.SetHasConstantBufferDrawParameters(true);
+
+            engine.UpdateState();
+
+            if (indexed)
+            {
+                _context.Renderer.Pipeline.DrawIndexed(count, instanceCount, firstIndex, firstVertex, firstInstance);
+                _state.State.FirstVertex = 0;
+            }
+            else
+            {
+                _context.Renderer.Pipeline.Draw(count, instanceCount, firstVertex, firstInstance);
+            }
+
+            _state.State.FirstInstance = 0;
+
+            _drawState.DrawIndexed = false;
+
+            if (renderEnable == ConditionalRenderEnabled.Host)
+            {
+                _context.Renderer.Pipeline.EndHostConditionalRendering();
+            }
+        }
+
+        /// <summary>
+        /// Performs a indirect draw, with parameters from a GPU buffer.
+        /// </summary>
+        /// <param name="engine">3D engine where this method is being called</param>
+        /// <param name="topology">Primitive topology</param>
+        /// <param name="indirectBufferAddress">Address of the buffer with the draw parameters, such as count, first index, etc</param>
+        /// <param name="parameterBufferAddress">Address of the buffer with the draw count</param>
+        /// <param name="maxDrawCount">Maximum number of draws that can be made</param>
+        /// <param name="stride">Distance in bytes between each entry on the data pointed to by <paramref name="indirectBufferAddress"/></param>
+        /// <param name="indexCount">Maximum number of indices that the draw can consume</param>
+        /// <param name="drawType">Type of the indirect draw, which can be indexed or non-indexed, with or without a draw count</param>
+        public void DrawIndirect(
+            ThreedClass engine,
+            PrimitiveTopology topology,
+            ulong indirectBufferAddress,
+            ulong parameterBufferAddress,
+            int maxDrawCount,
+            int stride,
+            int indexCount,
+            IndirectDrawType drawType)
+        {
+            UpdateTopology(topology);
+
+            ConditionalRenderEnabled renderEnable = ConditionalRendering.GetRenderEnable(
+                _context,
+                _channel.MemoryManager,
+                _state.State.RenderEnableAddress,
+                _state.State.RenderEnableCondition);
+
+            if (renderEnable == ConditionalRenderEnabled.False)
+            {
+                _drawState.DrawIndexed = false;
+                return;
+            }
+
+            PhysicalMemory memory = _channel.MemoryManager.Physical;
+
+            bool hasCount = (drawType & IndirectDrawType.Count) != 0;
+            bool indexed = (drawType & IndirectDrawType.Indexed) != 0;
+
+            if (indexed)
+            {
+                indexCount = Math.Clamp(indexCount, MinIndirectIndexCount, MaxIndirectIndexCount);
+                _drawState.FirstIndex = 0;
+                _drawState.IndexCount = indexCount;
+                engine.ForceStateDirty(IndexBufferCountMethodOffset * 4);
+            }
+
+            _drawState.DrawIndexed = indexed;
+            _drawState.DrawIndirect = true;
+            _currentSpecState.SetHasConstantBufferDrawParameters(true);
+
+            engine.UpdateState();
+
+            if (hasCount)
+            {
+                var indirectBuffer = memory.BufferCache.GetBufferRange(indirectBufferAddress, (ulong)maxDrawCount * (ulong)stride);
+                var parameterBuffer = memory.BufferCache.GetBufferRange(parameterBufferAddress, 4);
+
+                if (indexed)
+                {
+                    _context.Renderer.Pipeline.DrawIndexedIndirectCount(indirectBuffer, parameterBuffer, maxDrawCount, stride);
+                }
+                else
+                {
+                    _context.Renderer.Pipeline.DrawIndirectCount(indirectBuffer, parameterBuffer, maxDrawCount, stride);
+                }
+            }
+            else
+            {
+                var indirectBuffer = memory.BufferCache.GetBufferRange(indirectBufferAddress, (ulong)stride);
+
+                if (indexed)
+                {
+                    _context.Renderer.Pipeline.DrawIndexedIndirect(indirectBuffer);
+                }
+                else
+                {
+                    _context.Renderer.Pipeline.DrawIndirect(indirectBuffer);
+                }
+            }
+
+            _drawState.DrawIndexed = false;
+            _drawState.DrawIndirect = false;
+
+            if (renderEnable == ConditionalRenderEnabled.Host)
+            {
+                _context.Renderer.Pipeline.EndHostConditionalRendering();
+            }
+        }
+
+        /// <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;
+
+                bool indexedInline = _instancedIndexedInline;
+
+                if (_instancedIndexed || indexedInline)
+                {
+                    if (indexedInline)
+                    {
+                        int inlineIndexCount = _drawState.IbStreamer.GetAndResetInlineIndexCount(_context.Renderer);
+                        BufferRange br = new BufferRange(_drawState.IbStreamer.GetInlineIndexBuffer(), 0, inlineIndexCount * 4);
+
+                        _channel.BufferManager.SetIndexBuffer(br, IndexType.UInt);
+                    }
+
+                    _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 can also be specified with 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)
+        {
+            Clear(engine, argument, 1);
+        }
+
+        /// <summary>
+        /// Clears the current color and depth-stencil buffers.
+        /// Which buffers should be cleared can also specified with the arguments.
+        /// </summary>
+        /// <param name="engine">3D engine where this method is being called</param>
+        /// <param name="argument">Method call argument</param>
+        /// <param name="layerCount">For array and 3D textures, indicates how many layers should be cleared</param>
+        public void Clear(ThreedClass engine, int argument, int layerCount)
+        {
+            ConditionalRenderEnabled renderEnable = ConditionalRendering.GetRenderEnable(
+                _context,
+                _channel.MemoryManager,
+                _state.State.RenderEnableAddress,
+                _state.State.RenderEnableCondition);
+
+            if (renderEnable == ConditionalRenderEnabled.False)
+            {
+                return;
+            }
+
+            bool clearDepth = (argument & 1) != 0;
+            bool clearStencil = (argument & 2) != 0;
+            uint componentMask = (uint)((argument >> 2) & 0xf);
+            int index = (argument >> 6) & 0xf;
+            int layer = (argument >> 10) & 0x3ff;
+
+            RenderTargetUpdateFlags updateFlags = RenderTargetUpdateFlags.SingleColor;
+
+            if (layer != 0 || layerCount > 1)
+            {
+                updateFlags |= RenderTargetUpdateFlags.Layered;
+            }
+
+            if (clearDepth || clearStencil)
+            {
+                updateFlags |= RenderTargetUpdateFlags.UpdateDepthStencil;
+            }
+
+            engine.UpdateRenderTargetState(updateFlags, singleUse: componentMask != 0 ? index : -1);
+
+            // If there is a mismatch on the host clip region and the one explicitly defined by the guest
+            // on the screen scissor state, then we need to force only one texture to be bound to avoid
+            // host clipping.
+            var screenScissorState = _state.State.ScreenScissorState;
+
+            // Must happen after UpdateRenderTargetState to have up-to-date clip region values.
+            bool clipMismatch = (screenScissorState.X | screenScissorState.Y) != 0 ||
+                                screenScissorState.Width != _channel.TextureManager.ClipRegionWidth ||
+                                screenScissorState.Height != _channel.TextureManager.ClipRegionHeight;
+
+            bool clearAffectedByStencilMask = (_state.State.ClearFlags & 1) != 0;
+            bool clearAffectedByScissor = (_state.State.ClearFlags & 0x100) != 0;
+            bool needsCustomScissor = !clearAffectedByScissor || clipMismatch;
+
+            // Scissor and rasterizer discard also affect clears.
+            ulong updateMask = 1UL << StateUpdater.RasterizerStateIndex;
+
+            if (!needsCustomScissor)
+            {
+                updateMask |= 1UL << StateUpdater.ScissorStateIndex;
+            }
+
+            engine.UpdateState(updateMask);
+
+            if (needsCustomScissor)
+            {
+                int scissorX = screenScissorState.X;
+                int scissorY = screenScissorState.Y;
+                int scissorW = screenScissorState.Width;
+                int scissorH = screenScissorState.Height;
+
+                if (clearAffectedByScissor && _state.State.ScissorState[0].Enable)
+                {
+                    ref var scissorState = ref _state.State.ScissorState[0];
+
+                    scissorX = Math.Max(scissorX, scissorState.X1);
+                    scissorY = Math.Max(scissorY, scissorState.Y1);
+                    scissorW = Math.Min(scissorW, scissorState.X2 - scissorState.X1);
+                    scissorH = Math.Min(scissorH, scissorState.Y2 - scissorState.Y1);
+                }
+
+                float scale = _channel.TextureManager.RenderTargetScale;
+                if (scale != 1f)
+                {
+                    scissorX = (int)(scissorX * scale);
+                    scissorY = (int)(scissorY * scale);
+                    scissorW = (int)MathF.Ceiling(scissorW * scale);
+                    scissorH = (int)MathF.Ceiling(scissorH * scale);
+                }
+
+                Span<Rectangle<int>> scissors = stackalloc Rectangle<int>[]
+                {
+                    new Rectangle<int>(scissorX, scissorY, scissorW, scissorH)
+                };
+
+                _context.Renderer.Pipeline.SetScissors(scissors);
+            }
+
+            _channel.TextureManager.UpdateRenderTargets();
+
+            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, layer, layerCount, componentMask, color);
+            }
+
+            if (clearDepth || clearStencil)
+            {
+                float depthValue = _state.State.ClearDepthValue;
+                int stencilValue = (int)_state.State.ClearStencilValue;
+
+                int stencilMask = 0;
+
+                if (clearStencil)
+                {
+                    stencilMask = clearAffectedByStencilMask ? _state.State.StencilTestState.FrontMask : 0xff;
+                }
+
+                if (clipMismatch)
+                {
+                    _channel.TextureManager.UpdateRenderTargetDepthStencil();
+                }
+
+                _context.Renderer.Pipeline.ClearRenderTargetDepthStencil(
+                    layer,
+                    layerCount,
+                    depthValue,
+                    clearDepth,
+                    stencilValue,
+                    stencilMask);
+            }
+
+            if (needsCustomScissor)
+            {
+                engine.UpdateScissorState();
+            }
+
+            engine.UpdateRenderTargetState(RenderTargetUpdateFlags.UpdateAll);
+
+            if (renderEnable == ConditionalRenderEnabled.Host)
+            {
+                _context.Renderer.Pipeline.EndHostConditionalRendering();
+            }
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawState.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawState.cs
new file mode 100644
index 00000000..42ec2442
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/DrawState.cs
@@ -0,0 +1,65 @@
+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>
+        /// First vertex used on non-indexed draws. This value is stored somewhere else on indexed draws.
+        /// </summary>
+        public int DrawFirstVertex;
+
+        /// <summary>
+        /// Vertex count used on non-indexed draws. Indexed draws have a index count instead.
+        /// </summary>
+        public int DrawVertexCount;
+
+        /// <summary>
+        /// Indicates if the next draw will be a indexed draw.
+        /// </summary>
+        public bool DrawIndexed;
+
+        /// <summary>
+        /// Indicates if the next draw will be a indirect draw.
+        /// </summary>
+        public bool DrawIndirect;
+
+        /// <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>
+        /// Indicates that the draw is writing the base vertex, base instance and draw index to Constant Buffer 0.
+        /// </summary>
+        public bool HasConstantBufferDrawParameters;
+
+        /// <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();
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/IbStreamer.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/IbStreamer.cs
new file mode 100644
index 00000000..80d8c00b
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/IbStreamer.cs
@@ -0,0 +1,194 @@
+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 const int BufferCapacity = 256; // Must be a power of 2.
+
+        private BufferHandle _inlineIndexBuffer;
+        private int _inlineIndexBufferSize;
+        private int _inlineIndexCount;
+        private uint[] _buffer;
+        private int _bufferOffset;
+
+        /// <summary>
+        /// Indicates if any index buffer data has been pushed.
+        /// </summary>
+        public bool HasInlineIndexData => _inlineIndexCount != 0;
+
+        /// <summary>
+        /// Total numbers of indices that have been pushed.
+        /// </summary>
+        public int InlineIndexCount => _inlineIndexCount;
+
+        /// <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>
+        /// <param name="renderer">Host renderer</param>
+        /// <returns>Inline index bufffer count</returns>
+        public int GetAndResetInlineIndexCount(IRenderer renderer)
+        {
+            UpdateRemaining(renderer);
+            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);
+
+            int offset = _inlineIndexCount;
+
+            PushData(renderer, offset, i0);
+            PushData(renderer, offset + 1, i1);
+            PushData(renderer, offset + 2, i2);
+            PushData(renderer, offset + 3, i3);
+
+            _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);
+
+            int offset = _inlineIndexCount;
+
+            PushData(renderer, offset, i0);
+            PushData(renderer, offset + 1, i1);
+
+            _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;
+
+            int offset = _inlineIndexCount++;
+
+            PushData(renderer, offset, i0);
+        }
+
+        /// <summary>
+        /// Pushes a 32-bit value to the index buffer.
+        /// </summary>
+        /// <param name="renderer">Host renderer</param>
+        /// <param name="offset">Offset where the data should be written, in 32-bit words</param>
+        /// <param name="value">Index value to be written</param>
+        private void PushData(IRenderer renderer, int offset, uint value)
+        {
+            if (_buffer == null)
+            {
+                _buffer = new uint[BufferCapacity];
+            }
+
+            // We upload data in chunks.
+            // If we are at the start of a chunk, then the buffer might be full,
+            // in that case we need to submit any existing data before overwriting the buffer.
+            int subOffset = offset & (BufferCapacity - 1);
+
+            if (subOffset == 0 && offset != 0)
+            {
+                int baseOffset = (offset - BufferCapacity) * sizeof(uint);
+                BufferHandle buffer = GetInlineIndexBuffer(renderer, baseOffset, BufferCapacity * sizeof(uint));
+                renderer.SetBufferData(buffer, baseOffset, MemoryMarshal.Cast<uint, byte>(_buffer));
+            }
+
+            _buffer[subOffset] = value;
+        }
+
+        /// <summary>
+        /// Makes sure that any pending data is submitted to the GPU before the index buffer is used.
+        /// </summary>
+        /// <param name="renderer">Host renderer</param>
+        private void UpdateRemaining(IRenderer renderer)
+        {
+            int offset = _inlineIndexCount;
+            if (offset == 0)
+            {
+                return;
+            }
+
+            int count = offset & (BufferCapacity - 1);
+            if (count == 0)
+            {
+                count = BufferCapacity;
+            }
+
+            int baseOffset = (offset - count) * sizeof(uint);
+            int length = count * sizeof(uint);
+            BufferHandle buffer = GetInlineIndexBuffer(renderer, baseOffset, length);
+            renderer.SetBufferData(buffer, baseOffset, MemoryMarshal.Cast<uint, byte>(_buffer).Slice(0, length));
+        }
+
+        /// <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>
+        /// <param name="length">Number of bytes that will be written</param>
+        /// <returns>Buffer handle</returns>
+        private BufferHandle GetInlineIndexBuffer(IRenderer renderer, int offset, int length)
+        {
+            // 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 + length + 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;
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/IndirectDrawType.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/IndirectDrawType.cs
new file mode 100644
index 00000000..d78aa498
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/IndirectDrawType.cs
@@ -0,0 +1,38 @@
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Indirect draw type, which can be indexed or non-indexed, with or without a draw count.
+    /// </summary>
+    enum IndirectDrawType
+    {
+        /// <summary>
+        /// Non-indexed draw without draw count.
+        /// </summary>
+        DrawIndirect = 0,
+
+        /// <summary>
+        /// Indexed draw without draw count.
+        /// </summary>
+        DrawIndexedIndirect = Indexed,
+
+        /// <summary>
+        /// Non-indexed draw with draw count.
+        /// </summary>
+        DrawIndirectCount = Count,
+
+        /// <summary>
+        /// Indexed draw with draw count.
+        /// </summary>
+        DrawIndexedIndirectCount = Indexed | Count,
+
+        /// <summary>
+        /// Indexed flag.
+        /// </summary>
+        Indexed = 1 << 0,
+
+        /// <summary>
+        /// Draw count flag.
+        /// </summary>
+        Count = 1 << 1
+    }
+}
\ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/RenderTargetUpdateFlags.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/RenderTargetUpdateFlags.cs
new file mode 100644
index 00000000..cf2e818c
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/RenderTargetUpdateFlags.cs
@@ -0,0 +1,41 @@
+using System;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Flags indicating how the render targets should be updated.
+    /// </summary>
+    [Flags]
+    enum RenderTargetUpdateFlags
+    {
+        /// <summary>
+        /// No flags.
+        /// </summary>
+        None = 0,
+
+        /// <summary>
+        /// Get render target index from the control register.
+        /// </summary>
+        UseControl = 1 << 0,
+
+        /// <summary>
+        /// Indicates that all render targets are 2D array textures.
+        /// </summary>
+        Layered = 1 << 1,
+
+        /// <summary>
+        /// Indicates that only a single color target will be used.
+        /// </summary>
+        SingleColor = 1 << 2,
+
+        /// <summary>
+        /// Indicates that the depth-stencil target will be used.
+        /// </summary>
+        UpdateDepthStencil = 1 << 3,
+
+        /// <summary>
+        /// Default update flags for draw.
+        /// </summary>
+        UpdateAll = UseControl | UpdateDepthStencil
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/SemaphoreUpdater.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/SemaphoreUpdater.cs
new file mode 100644
index 00000000..63a2c841
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/SemaphoreUpdater.cs
@@ -0,0 +1,190 @@
+using Ryujinx.Graphics.GAL;
+using System;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Semaphore updater.
+    /// </summary>
+    class SemaphoreUpdater
+    {
+        /// <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
+        {
+            Payload = 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 = _context.GetTimestamp();
+
+            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.Payload:
+                    resultHandler(null, (ulong)_state.State.SemaphorePayload);
+                    break;
+                case ReportCounterType.SamplesPassed:
+                    counter = _context.Renderer.ReportCounter(CounterType.SamplesPassed, resultHandler, false);
+                    break;
+                case ReportCounterType.PrimitivesGenerated:
+                    counter = _context.Renderer.ReportCounter(CounterType.PrimitivesGenerated, resultHandler, false);
+                    break;
+                case ReportCounterType.TransformFeedbackPrimitivesWritten:
+                    counter = _context.Renderer.ReportCounter(CounterType.TransformFeedbackPrimitivesWritten, resultHandler, false);
+                    break;
+            }
+
+            _channel.MemoryManager.CounterCache.AddOrUpdate(gpuVa, counter);
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/SpecializationStateUpdater.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/SpecializationStateUpdater.cs
new file mode 100644
index 00000000..a8af5497
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/SpecializationStateUpdater.cs
@@ -0,0 +1,346 @@
+using Ryujinx.Common.Memory;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.Types;
+using Ryujinx.Graphics.Gpu.Shader;
+using Ryujinx.Graphics.Shader;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Maintains a "current" specialiation state, and provides a flag to check if it has changed meaningfully.
+    /// </summary>
+    internal class SpecializationStateUpdater
+    {
+        private readonly GpuContext _context;
+        private GpuChannelGraphicsState _graphics;
+        private GpuChannelPoolState _pool;
+
+        private bool _usesDrawParameters;
+        private bool _usesTopology;
+
+        private bool _changed;
+
+        /// <summary>
+        /// Creates a new instance of the specialization state updater class.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        public SpecializationStateUpdater(GpuContext context)
+        {
+            _context = context;
+        }
+
+        /// <summary>
+        /// Signal that the specialization state has changed.
+        /// </summary>
+        private void Signal()
+        {
+            _changed = true;
+        }
+
+        /// <summary>
+        /// Checks if the specialization state has changed since the last check.
+        /// </summary>
+        /// <returns>True if it has changed, false otherwise</returns>
+        public bool HasChanged()
+        {
+            if (_changed)
+            {
+                _changed = false;
+                return true;
+            }
+            else
+            {
+                return false;
+            }
+        }
+
+        /// <summary>
+        /// Sets the active shader, clearing the dirty state and recording if certain specializations are noteworthy.
+        /// </summary>
+        /// <param name="gs">The active shader</param>
+        public void SetShader(CachedShaderProgram gs)
+        {
+            _usesDrawParameters = gs.Shaders[1]?.Info.UsesDrawParameters ?? false;
+            _usesTopology = gs.SpecializationState.IsPrimitiveTopologyQueried();
+
+            _changed = false;
+        }
+
+        /// <summary>
+        /// Get the current graphics state.
+        /// </summary>
+        /// <returns>GPU graphics state</returns>
+        public ref GpuChannelGraphicsState GetGraphicsState()
+        {
+            return ref _graphics;
+        }
+
+        /// <summary>
+        /// Get the current pool state.
+        /// </summary>
+        /// <returns>GPU pool state</returns>
+        public ref GpuChannelPoolState GetPoolState()
+        {
+            return ref _pool;
+        }
+
+        /// <summary>
+        /// Early Z force enable.
+        /// </summary>
+        /// <param name="value">The new value</param>
+        public void SetEarlyZForce(bool value)
+        {
+            _graphics.EarlyZForce = value;
+
+            Signal();
+        }
+
+        /// <summary>
+        /// Primitive topology of current draw.
+        /// </summary>
+        /// <param name="value">The new value</param>
+        public void SetTopology(PrimitiveTopology value)
+        {
+            if (value != _graphics.Topology)
+            {
+                _graphics.Topology = value;
+
+                if (_usesTopology)
+                {
+                    Signal();
+                }
+            }
+        }
+
+        /// <summary>
+        /// Tessellation mode.
+        /// </summary>
+        /// <param name="value">The new value</param>
+        public void SetTessellationMode(TessMode value)
+        {
+            if (value.Packed != _graphics.TessellationMode.Packed)
+            {
+                _graphics.TessellationMode = value;
+
+                Signal();
+            }
+        }
+
+        /// <summary>
+        /// Updates alpha-to-coverage state, and sets it as changed.
+        /// </summary>
+        /// <param name="enable">Whether alpha-to-coverage is enabled</param>
+        /// <param name="ditherEnable">Whether alpha-to-coverage dithering is enabled</param>
+        public void SetAlphaToCoverageEnable(bool enable, bool ditherEnable)
+        {
+            _graphics.AlphaToCoverageEnable = enable;
+            _graphics.AlphaToCoverageDitherEnable = ditherEnable;
+
+            Signal();
+        }
+
+        /// <summary>
+        /// Indicates whether the viewport transform is disabled.
+        /// </summary>
+        /// <param name="value">The new value</param>
+        public void SetViewportTransformDisable(bool value)
+        {
+            if (value != _graphics.ViewportTransformDisable)
+            {
+                _graphics.ViewportTransformDisable = value;
+
+                Signal();
+            }
+        }
+
+        /// <summary>
+        /// Depth mode zero to one or minus one to one.
+        /// </summary>
+        /// <param name="value">The new value</param>
+        public void SetDepthMode(bool value)
+        {
+            if (value != _graphics.DepthMode)
+            {
+                _graphics.DepthMode = value;
+
+                Signal();
+            }
+        }
+
+        /// <summary>
+        /// Indicates if the point size is set on the shader or is fixed.
+        /// </summary>
+        /// <param name="value">The new value</param>
+        public void SetProgramPointSizeEnable(bool value)
+        {
+            if (value != _graphics.ProgramPointSizeEnable)
+            {
+                _graphics.ProgramPointSizeEnable = value;
+
+                Signal();
+            }
+        }
+
+        /// <summary>
+        /// Point size used if <see cref="SetProgramPointSizeEnable" /> is provided false.
+        /// </summary>
+        /// <param name="value">The new value</param>
+        public void SetPointSize(float value)
+        {
+            if (value != _graphics.PointSize)
+            {
+                _graphics.PointSize = value;
+
+                Signal();
+            }
+        }
+
+        /// <summary>
+        /// Updates alpha test specialization state, and sets it as changed.
+        /// </summary>
+        /// <param name="enable">Whether alpha test is enabled</param>
+        /// <param name="reference">The value to compare with the fragment output alpha</param>
+        /// <param name="op">The comparison that decides if the fragment should be discarded</param>
+        public void SetAlphaTest(bool enable, float reference, CompareOp op)
+        {
+            _graphics.AlphaTestEnable = enable;
+            _graphics.AlphaTestReference = reference;
+            _graphics.AlphaTestCompare = op;
+
+            Signal();
+        }
+
+        /// <summary>
+        /// Updates the type of the vertex attributes consumed by the shader.
+        /// </summary>
+        /// <param name="state">The new state</param>
+        public void SetAttributeTypes(ref Array32<VertexAttribState> state)
+        {
+            bool changed = false;
+            ref Array32<AttributeType> attributeTypes = ref _graphics.AttributeTypes;
+
+            for (int location = 0; location < state.Length; location++)
+            {
+                VertexAttribType type = state[location].UnpackType();
+
+                AttributeType value = type switch
+                {
+                    VertexAttribType.Sint => AttributeType.Sint,
+                    VertexAttribType.Uint => AttributeType.Uint,
+                    _ => AttributeType.Float
+                };
+
+                if (attributeTypes[location] != value)
+                {
+                    attributeTypes[location] = value;
+                    changed = true;
+                }
+            }
+
+            if (changed)
+            {
+                Signal();
+            }
+        }
+
+        /// <summary>
+        /// Updates the type of the outputs produced by the fragment shader based on the current render target state.
+        /// </summary>
+        /// <param name="rtControl">The render target control register</param>
+        /// <param name="state">The color attachment state</param>
+        public void SetFragmentOutputTypes(RtControl rtControl, ref Array8<RtColorState> state)
+        {
+            bool changed = false;
+            int count = rtControl.UnpackCount();
+
+            for (int index = 0; index < Constants.TotalRenderTargets; index++)
+            {
+                int rtIndex = rtControl.UnpackPermutationIndex(index);
+
+                var colorState = state[rtIndex];
+
+                if (index < count && StateUpdater.IsRtEnabled(colorState))
+                {
+                    Format format = colorState.Format.Convert().Format;
+
+                    AttributeType type = format.IsInteger() ? (format.IsSint() ? AttributeType.Sint : AttributeType.Uint) : AttributeType.Float;
+
+                    if (type != _graphics.FragmentOutputTypes[index])
+                    {
+                        _graphics.FragmentOutputTypes[index] = type;
+                        changed = true;
+                    }
+                }
+            }
+
+            if (changed && _context.Capabilities.NeedsFragmentOutputSpecialization)
+            {
+                Signal();
+            }
+        }
+
+        /// <summary>
+        /// Indicates that the draw is writing the base vertex, base instance and draw index to Constant Buffer 0.
+        /// </summary>
+        /// <param name="value">The new value</param>
+        public void SetHasConstantBufferDrawParameters(bool value)
+        {
+            if (value != _graphics.HasConstantBufferDrawParameters)
+            {
+                _graphics.HasConstantBufferDrawParameters = value;
+
+                if (_usesDrawParameters)
+                {
+                    Signal();
+                }
+            }
+        }
+
+        /// <summary>
+        /// Indicates that any storage buffer use is unaligned.
+        /// </summary>
+        /// <param name="value">The new value</param>
+        /// <returns>True if the unaligned state changed, false otherwise</returns>
+        public bool SetHasUnalignedStorageBuffer(bool value)
+        {
+            if (value != _graphics.HasUnalignedStorageBuffer)
+            {
+                _graphics.HasUnalignedStorageBuffer = value;
+
+                Signal();
+
+                return true;
+            }
+
+            return false;
+        }
+
+        /// <summary>
+        /// Sets the GPU pool state.
+        /// </summary>
+        /// <param name="state">The new state</param>
+        public void SetPoolState(GpuChannelPoolState state)
+        {
+            if (!state.Equals(_pool))
+            {
+                _pool = state;
+
+                Signal();
+            }
+        }
+
+        /// <summary>
+        /// Sets the dual-source blend enabled state.
+        /// </summary>
+        /// <param name="enabled">True if blending is enabled and using dual-source blend</param>
+        public void SetDualSourceBlendEnabled(bool enabled)
+        {
+            if (enabled != _graphics.DualSourceBlendEnable)
+            {
+                _graphics.DualSourceBlendEnable = enabled;
+
+                Signal();
+            }
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdateTracker.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdateTracker.cs
new file mode 100644
index 00000000..7c730967
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdateTracker.cs
@@ -0,0 +1,177 @@
+using Ryujinx.Graphics.Device;
+using System;
+using System.Collections.Generic;
+using System.Diagnostics;
+using System.Diagnostics.CodeAnalysis;
+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>
+    readonly 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<[DynamicallyAccessedMembers(DynamicallyAccessedMemberTypes.PublicFields)] 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 if the given register group is dirty without clearing it.
+        /// </summary>
+        /// <param name="groupIndex">Index of the group to check</param>
+        /// <returns>True if dirty, false otherwise</returns>
+        public bool IsDirty(int groupIndex)
+        {
+            return (_dirtyMask & (1UL << groupIndex)) != 0;
+        }
+
+        /// <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;
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdater.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdater.cs
new file mode 100644
index 00000000..00e09a31
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/StateUpdater.cs
@@ -0,0 +1,1448 @@
+using Ryujinx.Common.Logging;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.Threed.Blender;
+using Ryujinx.Graphics.Gpu.Engine.Types;
+using Ryujinx.Graphics.Gpu.Image;
+using Ryujinx.Graphics.Gpu.Shader;
+using Ryujinx.Graphics.Shader;
+using Ryujinx.Graphics.Texture;
+using System;
+using System.Runtime.CompilerServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// GPU state updater.
+    /// </summary>
+    class StateUpdater
+    {
+        public const int ShaderStateIndex = 26;
+        public const int RasterizerStateIndex = 15;
+        public const int ScissorStateIndex = 16;
+        public const int VertexBufferStateIndex = 0;
+        public const int PrimitiveRestartStateIndex = 12;
+        public const int RenderTargetStateIndex = 27;
+
+        private readonly GpuContext _context;
+        private readonly GpuChannel _channel;
+        private readonly DeviceStateWithShadow<ThreedClassState> _state;
+        private readonly DrawState _drawState;
+        private readonly AdvancedBlendManager _blendManager;
+
+        private readonly StateUpdateTracker<ThreedClassState> _updateTracker;
+
+        private readonly ShaderProgramInfo[] _currentProgramInfo;
+        private ShaderSpecializationState _shaderSpecState;
+        private SpecializationStateUpdater _currentSpecState;
+
+        private ProgramPipelineState _pipeline;
+
+        private bool _vsUsesDrawParameters;
+        private bool _vtgWritesRtLayer;
+        private byte _vsClipDistancesWritten;
+        private uint _vbEnableMask;
+
+        private bool _prevDrawIndexed;
+        private bool _prevDrawIndirect;
+        private IndexType _prevIndexType;
+        private uint _prevFirstVertex;
+        private bool _prevTfEnable;
+
+        private uint _prevRtNoAlphaMask;
+
+        /// <summary>
+        /// Creates a new instance of the state updater.
+        /// </summary>
+        /// <param name="context">GPU context</param>
+        /// <param name="channel">GPU channel</param>
+        /// <param name="state">3D engine state</param>
+        /// <param name="drawState">Draw state</param>
+        /// <param name="blendManager">Advanced blend manager</param>
+        /// <param name="spec">Specialization state updater</param>
+        public StateUpdater(
+            GpuContext context,
+            GpuChannel channel,
+            DeviceStateWithShadow<ThreedClassState> state,
+            DrawState drawState,
+            AdvancedBlendManager blendManager,
+            SpecializationStateUpdater spec)
+        {
+            _context = context;
+            _channel = channel;
+            _state = state;
+            _drawState = drawState;
+            _blendManager = blendManager;
+            _currentProgramInfo = new ShaderProgramInfo[Constants.ShaderStages];
+            _currentSpecState = spec;
+
+            // ShaderState must be updated after other state updates, as specialization/pipeline state is used when fetching shaders.
+            // Render target state must appear after shader state as it depends on information from the currently bound shader.
+            // Rasterizer and scissor states are checked by render target clear, their indexes
+            // must be updated on the constants "RasterizerStateIndex" and "ScissorStateIndex" if modified.
+            // The vertex buffer state may be forced dirty when a indexed draw starts, the "VertexBufferStateIndex"
+            // constant must be updated if modified.
+            // The order of the other state updates doesn't matter.
+            _updateTracker = new StateUpdateTracker<ThreedClassState>(new[]
+            {
+                new StateUpdateCallbackEntry(UpdateVertexBufferState,
+                    nameof(ThreedClassState.VertexBufferDrawState),
+                    nameof(ThreedClassState.VertexBufferInstanced),
+                    nameof(ThreedClassState.VertexBufferState),
+                    nameof(ThreedClassState.VertexBufferEndAddress)),
+
+                // Must be done after vertex buffer updates.
+                new StateUpdateCallbackEntry(UpdateVertexAttribState, nameof(ThreedClassState.VertexAttribState)),
+
+                new StateUpdateCallbackEntry(UpdateBlendState,
+                    nameof(ThreedClassState.BlendUcodeEnable),
+                    nameof(ThreedClassState.BlendUcodeSize),
+                    nameof(ThreedClassState.BlendIndependent),
+                    nameof(ThreedClassState.BlendConstant),
+                    nameof(ThreedClassState.BlendStateCommon),
+                    nameof(ThreedClassState.BlendEnableCommon),
+                    nameof(ThreedClassState.BlendEnable),
+                    nameof(ThreedClassState.BlendState)),
+
+                new StateUpdateCallbackEntry(UpdateFaceState, nameof(ThreedClassState.FaceState)),
+
+                new StateUpdateCallbackEntry(UpdateStencilTestState,
+                    nameof(ThreedClassState.StencilBackMasks),
+                    nameof(ThreedClassState.StencilTestState),
+                    nameof(ThreedClassState.StencilBackTestState)),
+
+                new StateUpdateCallbackEntry(UpdateDepthTestState,
+                    nameof(ThreedClassState.DepthTestEnable),
+                    nameof(ThreedClassState.DepthWriteEnable),
+                    nameof(ThreedClassState.DepthTestFunc)),
+
+                new StateUpdateCallbackEntry(UpdateTessellationState,
+                    nameof(ThreedClassState.TessMode),
+                    nameof(ThreedClassState.TessOuterLevel),
+                    nameof(ThreedClassState.TessInnerLevel),
+                    nameof(ThreedClassState.PatchVertices)),
+
+                new StateUpdateCallbackEntry(UpdateViewportTransform,
+                    nameof(ThreedClassState.DepthMode),
+                    nameof(ThreedClassState.ViewportTransform),
+                    nameof(ThreedClassState.ViewportExtents),
+                    nameof(ThreedClassState.YControl),
+                    nameof(ThreedClassState.ViewportTransformEnable)),
+
+                new StateUpdateCallbackEntry(UpdateLogicOpState, nameof(ThreedClassState.LogicOpState)),
+
+                new StateUpdateCallbackEntry(UpdateDepthClampState, nameof(ThreedClassState.ViewVolumeClipControl)),
+
+                new StateUpdateCallbackEntry(UpdatePolygonMode,
+                    nameof(ThreedClassState.PolygonModeFront),
+                    nameof(ThreedClassState.PolygonModeBack)),
+
+                new StateUpdateCallbackEntry(UpdateDepthBiasState,
+                    nameof(ThreedClassState.DepthBiasState),
+                    nameof(ThreedClassState.DepthBiasFactor),
+                    nameof(ThreedClassState.DepthBiasUnits),
+                    nameof(ThreedClassState.DepthBiasClamp)),
+
+                new StateUpdateCallbackEntry(UpdatePrimitiveRestartState, nameof(ThreedClassState.PrimitiveRestartState)),
+
+                new StateUpdateCallbackEntry(UpdateLineState,
+                    nameof(ThreedClassState.LineWidthSmooth),
+                    nameof(ThreedClassState.LineSmoothEnable)),
+
+                new StateUpdateCallbackEntry(UpdateRtColorMask,
+                    nameof(ThreedClassState.RtColorMaskShared),
+                    nameof(ThreedClassState.RtColorMask)),
+
+                new StateUpdateCallbackEntry(UpdateRasterizerState, nameof(ThreedClassState.RasterizeEnable)),
+
+                new StateUpdateCallbackEntry(UpdateScissorState,
+                    nameof(ThreedClassState.ScissorState),
+                    nameof(ThreedClassState.ScreenScissorState)),
+
+                new StateUpdateCallbackEntry(UpdateTfBufferState, nameof(ThreedClassState.TfBufferState)),
+                new StateUpdateCallbackEntry(UpdateUserClipState, nameof(ThreedClassState.ClipDistanceEnable)),
+
+                new StateUpdateCallbackEntry(UpdateAlphaTestState,
+                    nameof(ThreedClassState.AlphaTestEnable),
+                    nameof(ThreedClassState.AlphaTestRef),
+                    nameof(ThreedClassState.AlphaTestFunc)),
+
+                new StateUpdateCallbackEntry(UpdateSamplerPoolState,
+                    nameof(ThreedClassState.SamplerPoolState),
+                    nameof(ThreedClassState.SamplerIndex)),
+
+                new StateUpdateCallbackEntry(UpdateTexturePoolState, nameof(ThreedClassState.TexturePoolState)),
+
+                new StateUpdateCallbackEntry(UpdatePointState,
+                    nameof(ThreedClassState.PointSize),
+                    nameof(ThreedClassState.VertexProgramPointSize),
+                    nameof(ThreedClassState.PointSpriteEnable),
+                    nameof(ThreedClassState.PointCoordReplace)),
+
+                new StateUpdateCallbackEntry(UpdateIndexBufferState,
+                    nameof(ThreedClassState.IndexBufferState),
+                    nameof(ThreedClassState.IndexBufferCount)),
+
+                new StateUpdateCallbackEntry(UpdateMultisampleState,
+                    nameof(ThreedClassState.AlphaToCoverageDitherEnable),
+                    nameof(ThreedClassState.MultisampleControl)),
+
+                new StateUpdateCallbackEntry(UpdateEarlyZState,
+                    nameof(ThreedClassState.EarlyZForce)),
+
+                new StateUpdateCallbackEntry(UpdateShaderState,
+                    nameof(ThreedClassState.ShaderBaseAddress),
+                    nameof(ThreedClassState.ShaderState)),
+
+                new StateUpdateCallbackEntry(UpdateRenderTargetState,
+                    nameof(ThreedClassState.RtColorState),
+                    nameof(ThreedClassState.RtDepthStencilState),
+                    nameof(ThreedClassState.RtControl),
+                    nameof(ThreedClassState.RtDepthStencilSize),
+                    nameof(ThreedClassState.RtDepthStencilEnable)),
+            });
+        }
+
+        /// <summary>
+        /// Sets a register at a specific offset as dirty.
+        /// This must be called if the register value was modified.
+        /// </summary>
+        /// <param name="offset">Register offset</param>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void SetDirty(int offset)
+        {
+            _updateTracker.SetDirty(offset);
+        }
+
+        /// <summary>
+        /// Force all the guest state to be marked as dirty.
+        /// The next call to <see cref="Update"/> will update all the host state.
+        /// </summary>
+        public void SetAllDirty()
+        {
+            _updateTracker.SetAllDirty();
+        }
+
+        /// <summary>
+        /// Updates host state for any modified guest state, since the last time this function was called.
+        /// </summary>
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        public void Update()
+        {
+            // The vertex buffer size is calculated using a different
+            // method when doing indexed draws, so we need to make sure
+            // to update the vertex buffers if we are doing a regular
+            // draw after a indexed one and vice-versa.
+            if (_drawState.DrawIndexed != _prevDrawIndexed)
+            {
+                _updateTracker.ForceDirty(VertexBufferStateIndex);
+
+                // If PrimitiveRestartDrawArrays is false and this is a non-indexed draw, we need to ensure primitive restart is disabled.
+                // If PrimitiveRestartDrawArrays is false and this is a indexed draw, we need to ensure primitive restart enable matches GPU state.
+                // If PrimitiveRestartDrawArrays is true, then primitive restart enable should always match GPU state.
+                // That is because "PrimitiveRestartDrawArrays" is not configurable on the backend, it is always
+                // true on OpenGL and always false on Vulkan.
+                if (!_state.State.PrimitiveRestartDrawArrays && _state.State.PrimitiveRestartState.Enable)
+                {
+                    _updateTracker.ForceDirty(PrimitiveRestartStateIndex);
+                }
+
+                _prevDrawIndexed = _drawState.DrawIndexed;
+            }
+
+            // Some draw parameters are used to restrict the vertex buffer size,
+            // but they can't be used on indirect draws because their values are unknown in this case.
+            // When switching between indirect and non-indirect draw, we need to
+            // make sure the vertex buffer sizes are still correct.
+            if (_drawState.DrawIndirect != _prevDrawIndirect)
+            {
+                _updateTracker.ForceDirty(VertexBufferStateIndex);
+            }
+
+            // In some cases, the index type is also used to guess the
+            // vertex buffer size, so we must update it if the type changed too.
+            if (_drawState.DrawIndexed &&
+                (_prevIndexType != _state.State.IndexBufferState.Type ||
+                 _prevFirstVertex != _state.State.FirstVertex))
+            {
+                _updateTracker.ForceDirty(VertexBufferStateIndex);
+                _prevIndexType = _state.State.IndexBufferState.Type;
+                _prevFirstVertex = _state.State.FirstVertex;
+            }
+
+            bool tfEnable = _state.State.TfEnable;
+
+            if (!tfEnable && _prevTfEnable)
+            {
+                _context.Renderer.Pipeline.EndTransformFeedback();
+                _prevTfEnable = false;
+            }
+
+            if (_updateTracker.IsDirty(RenderTargetStateIndex))
+            {
+                UpdateRenderTargetSpecialization();
+            }
+
+            _updateTracker.Update(ulong.MaxValue);
+
+            // If any state that the shader depends on changed,
+            // then we may need to compile/bind a different version
+            // of the shader for the new state.
+            if (_shaderSpecState != null && _currentSpecState.HasChanged())
+            {
+                if (!_shaderSpecState.MatchesGraphics(_channel, ref _currentSpecState.GetPoolState(), ref _currentSpecState.GetGraphicsState(), _vsUsesDrawParameters, false))
+                {
+                    // Shader must be reloaded. _vtgWritesRtLayer should not change.
+                    UpdateShaderState();
+                }
+            }
+
+            CommitBindings();
+
+            if (tfEnable && !_prevTfEnable)
+            {
+                _context.Renderer.Pipeline.BeginTransformFeedback(_drawState.Topology);
+                _prevTfEnable = true;
+            }
+        }
+
+        /// <summary>
+        /// Updates the host state for any modified guest state group with the respective bit set on <paramref name="mask"/>.
+        /// </summary>
+        /// <param name="mask">Mask, where each bit set corresponds to a group index that should be checked and updated</param>
+        public void Update(ulong mask)
+        {
+            _updateTracker.Update(mask);
+        }
+
+        /// <summary>
+        /// Ensures that the bindings are visible to the host GPU.
+        /// Note: this actually performs the binding using the host graphics API.
+        /// </summary>
+        private void CommitBindings()
+        {
+            UpdateStorageBuffers();
+
+            bool unalignedChanged = _currentSpecState.SetHasUnalignedStorageBuffer(_channel.BufferManager.HasUnalignedStorageBuffers);
+
+            if (!_channel.TextureManager.CommitGraphicsBindings(_shaderSpecState) || unalignedChanged)
+            {
+                // Shader must be reloaded. _vtgWritesRtLayer should not change.
+                UpdateShaderState();
+            }
+
+            _channel.BufferManager.CommitGraphicsBindings();
+        }
+
+        /// <summary>
+        /// Updates storage buffer bindings.
+        /// </summary>
+        private void UpdateStorageBuffers()
+        {
+            for (int stage = 0; stage < Constants.ShaderStages; stage++)
+            {
+                ShaderProgramInfo info = _currentProgramInfo[stage];
+
+                if (info == null)
+                {
+                    continue;
+                }
+
+                for (int index = 0; index < info.SBuffers.Count; index++)
+                {
+                    BufferDescriptor sb = info.SBuffers[index];
+
+                    ulong sbDescAddress = _channel.BufferManager.GetGraphicsUniformBufferAddress(stage, 0);
+
+                    int sbDescOffset = 0x110 + stage * 0x100 + sb.Slot * 0x10;
+
+                    sbDescAddress += (ulong)sbDescOffset;
+
+                    SbDescriptor sbDescriptor = _channel.MemoryManager.Physical.Read<SbDescriptor>(sbDescAddress);
+
+                    _channel.BufferManager.SetGraphicsStorageBuffer(stage, sb.Slot, sbDescriptor.PackAddress(), (uint)sbDescriptor.Size, sb.Flags);
+                }
+            }
+        }
+
+        /// <summary>
+        /// Updates tessellation state based on the guest GPU state.
+        /// </summary>
+        private void UpdateTessellationState()
+        {
+            _pipeline.PatchControlPoints = (uint)_state.State.PatchVertices;
+
+            _context.Renderer.Pipeline.SetPatchParameters(
+                _state.State.PatchVertices,
+                _state.State.TessOuterLevel.AsSpan(),
+                _state.State.TessInnerLevel.AsSpan());
+
+            _currentSpecState.SetTessellationMode(_state.State.TessMode);
+        }
+
+        /// <summary>
+        /// Updates transform feedback buffer state based on the guest GPU state.
+        /// </summary>
+        private void UpdateTfBufferState()
+        {
+            for (int index = 0; index < Constants.TotalTransformFeedbackBuffers; index++)
+            {
+                TfBufferState tfb = _state.State.TfBufferState[index];
+
+                if (!tfb.Enable)
+                {
+                    _channel.BufferManager.SetTransformFeedbackBuffer(index, 0, 0);
+
+                    continue;
+                }
+
+                _channel.BufferManager.SetTransformFeedbackBuffer(index, tfb.Address.Pack(), (uint)tfb.Size);
+            }
+        }
+
+        /// <summary>
+        /// Updates Rasterizer primitive discard state based on guest gpu state.
+        /// </summary>
+        private void UpdateRasterizerState()
+        {
+            bool enable = _state.State.RasterizeEnable;
+            _pipeline.RasterizerDiscard = !enable;
+            _context.Renderer.Pipeline.SetRasterizerDiscard(!enable);
+        }
+
+        /// <summary>
+        /// Updates render targets (color and depth-stencil buffers) based on current render target state.
+        /// </summary>
+        private void UpdateRenderTargetState()
+        {
+            UpdateRenderTargetState(RenderTargetUpdateFlags.UpdateAll);
+        }
+
+        /// <summary>
+        /// Updates render targets (color and depth-stencil buffers) based on current render target state.
+        /// </summary>
+        /// <param name="updateFlags">Flags indicating which render targets should be updated and how</param>
+        /// <param name="singleUse">If this is not -1, it indicates that only the given indexed target will be used.</param>
+        public void UpdateRenderTargetState(RenderTargetUpdateFlags updateFlags, int singleUse = -1)
+        {
+            var memoryManager = _channel.MemoryManager;
+            var rtControl = _state.State.RtControl;
+
+            bool useControl = updateFlags.HasFlag(RenderTargetUpdateFlags.UseControl);
+            bool layered = updateFlags.HasFlag(RenderTargetUpdateFlags.Layered);
+            bool singleColor = updateFlags.HasFlag(RenderTargetUpdateFlags.SingleColor);
+
+            int count = useControl ? rtControl.UnpackCount() : Constants.TotalRenderTargets;
+
+            var msaaMode = _state.State.RtMsaaMode;
+
+            int samplesInX = msaaMode.SamplesInX();
+            int samplesInY = msaaMode.SamplesInY();
+
+            var scissor = _state.State.ScreenScissorState;
+            Size sizeHint = new Size((scissor.X + scissor.Width) * samplesInX, (scissor.Y + scissor.Height) * samplesInY, 1);
+
+            int clipRegionWidth = int.MaxValue;
+            int clipRegionHeight = int.MaxValue;
+
+            bool changedScale = false;
+            uint rtNoAlphaMask = 0;
+
+            for (int index = 0; index < Constants.TotalRenderTargets; index++)
+            {
+                int rtIndex = useControl ? rtControl.UnpackPermutationIndex(index) : index;
+
+                var colorState = _state.State.RtColorState[rtIndex];
+
+                if (index >= count || !IsRtEnabled(colorState) || (singleColor && index != singleUse))
+                {
+                    changedScale |= _channel.TextureManager.SetRenderTargetColor(index, null);
+
+                    continue;
+                }
+
+                if (colorState.Format.NoAlpha())
+                {
+                    rtNoAlphaMask |= 1u << index;
+                }
+
+                Image.Texture color = memoryManager.Physical.TextureCache.FindOrCreateTexture(
+                    memoryManager,
+                    colorState,
+                    _vtgWritesRtLayer || layered,
+                    samplesInX,
+                    samplesInY,
+                    sizeHint);
+
+                changedScale |= _channel.TextureManager.SetRenderTargetColor(index, color);
+
+                if (color != null)
+                {
+                    if (clipRegionWidth > color.Width / samplesInX)
+                    {
+                        clipRegionWidth = color.Width / samplesInX;
+                    }
+
+                    if (clipRegionHeight > color.Height / samplesInY)
+                    {
+                        clipRegionHeight = color.Height / samplesInY;
+                    }
+                }
+            }
+
+            bool dsEnable = _state.State.RtDepthStencilEnable;
+
+            Image.Texture depthStencil = null;
+
+            if (dsEnable && updateFlags.HasFlag(RenderTargetUpdateFlags.UpdateDepthStencil))
+            {
+                var dsState = _state.State.RtDepthStencilState;
+                var dsSize = _state.State.RtDepthStencilSize;
+
+                depthStencil = memoryManager.Physical.TextureCache.FindOrCreateTexture(
+                    memoryManager,
+                    dsState,
+                    dsSize,
+                    _vtgWritesRtLayer || layered,
+                    samplesInX,
+                    samplesInY,
+                    sizeHint);
+
+                if (depthStencil != null)
+                {
+                    if (clipRegionWidth > depthStencil.Width / samplesInX)
+                    {
+                        clipRegionWidth = depthStencil.Width / samplesInX;
+                    }
+
+                    if (clipRegionHeight > depthStencil.Height / samplesInY)
+                    {
+                        clipRegionHeight = depthStencil.Height / samplesInY;
+                    }
+                }
+            }
+
+            changedScale |= _channel.TextureManager.SetRenderTargetDepthStencil(depthStencil);
+
+            if (changedScale)
+            {
+                float oldScale = _channel.TextureManager.RenderTargetScale;
+                _channel.TextureManager.UpdateRenderTargetScale(singleUse);
+
+                if (oldScale != _channel.TextureManager.RenderTargetScale)
+                {
+                    _context.Renderer.Pipeline.SetRenderTargetScale(_channel.TextureManager.RenderTargetScale);
+
+                    UpdateViewportTransform();
+                    UpdateScissorState();
+                }
+            }
+
+            _channel.TextureManager.SetClipRegion(clipRegionWidth, clipRegionHeight);
+
+            if (useControl && _prevRtNoAlphaMask != rtNoAlphaMask)
+            {
+                _prevRtNoAlphaMask = rtNoAlphaMask;
+
+                UpdateBlendState();
+            }
+        }
+
+        /// <summary>
+        /// Updates specialization state based on render target state.
+        /// </summary>
+        public void UpdateRenderTargetSpecialization()
+        {
+            _currentSpecState.SetFragmentOutputTypes(_state.State.RtControl, ref _state.State.RtColorState);
+        }
+
+        /// <summary>
+        /// Checks if a render target color buffer is used.
+        /// </summary>
+        /// <param name="colorState">Color buffer information</param>
+        /// <returns>True if the specified buffer is enabled/used, false otherwise</returns>
+        internal static bool IsRtEnabled(RtColorState colorState)
+        {
+            // Colors are disabled by writing 0 to the format.
+            return colorState.Format != 0 && colorState.WidthOrStride != 0;
+        }
+
+        /// <summary>
+        /// Updates host scissor test state based on current GPU state.
+        /// </summary>
+        public void UpdateScissorState()
+        {
+            const int MinX = 0;
+            const int MinY = 0;
+            const int MaxW = 0xffff;
+            const int MaxH = 0xffff;
+
+            Span<Rectangle<int>> regions = stackalloc Rectangle<int>[Constants.TotalViewports];
+
+            for (int index = 0; index < Constants.TotalViewports; index++)
+            {
+                ScissorState scissor = _state.State.ScissorState[index];
+
+                bool enable = scissor.Enable && (scissor.X1 != MinX ||
+                                                 scissor.Y1 != MinY ||
+                                                 scissor.X2 != MaxW ||
+                                                 scissor.Y2 != MaxH);
+
+                if (enable)
+                {
+                    int x = scissor.X1;
+                    int y = scissor.Y1;
+                    int width = scissor.X2 - x;
+                    int height = scissor.Y2 - y;
+
+                    if (_state.State.YControl.HasFlag(YControl.NegateY))
+                    {
+                        ref var screenScissor = ref _state.State.ScreenScissorState;
+                        y = screenScissor.Height - height - y;
+
+                        if (y < 0)
+                        {
+                            height += y;
+                            y = 0;
+                        }
+                    }
+
+                    float scale = _channel.TextureManager.RenderTargetScale;
+                    if (scale != 1f)
+                    {
+                        x = (int)(x * scale);
+                        y = (int)(y * scale);
+                        width = (int)MathF.Ceiling(width * scale);
+                        height = (int)MathF.Ceiling(height * scale);
+                    }
+
+                    regions[index] = new Rectangle<int>(x, y, width, height);
+                }
+                else
+                {
+                    regions[index] = new Rectangle<int>(MinX, MinY, MaxW, MaxH);
+                }
+            }
+
+            _context.Renderer.Pipeline.SetScissors(regions);
+        }
+
+        /// <summary>
+        /// Updates host depth clamp state based on current GPU state.
+        /// </summary>
+        /// <param name="state">Current GPU state</param>
+        private void UpdateDepthClampState()
+        {
+            ViewVolumeClipControl clip = _state.State.ViewVolumeClipControl;
+            bool clamp = (clip & ViewVolumeClipControl.DepthClampDisabled) == 0;
+
+            _pipeline.DepthClampEnable = clamp;
+            _context.Renderer.Pipeline.SetDepthClamp(clamp);
+        }
+
+        /// <summary>
+        /// Updates host alpha test state based on current GPU state.
+        /// </summary>
+        private void UpdateAlphaTestState()
+        {
+            _context.Renderer.Pipeline.SetAlphaTest(
+                _state.State.AlphaTestEnable,
+                _state.State.AlphaTestRef,
+                _state.State.AlphaTestFunc);
+
+            _currentSpecState.SetAlphaTest(
+                _state.State.AlphaTestEnable,
+                _state.State.AlphaTestRef,
+                _state.State.AlphaTestFunc);
+        }
+
+        /// <summary>
+        /// Updates host depth test state based on current GPU state.
+        /// </summary>
+        private void UpdateDepthTestState()
+        {
+            DepthTestDescriptor descriptor = new DepthTestDescriptor(
+                _state.State.DepthTestEnable,
+                _state.State.DepthWriteEnable,
+                _state.State.DepthTestFunc);
+
+            _pipeline.DepthTest = descriptor;
+            _context.Renderer.Pipeline.SetDepthTest(descriptor);
+        }
+
+        /// <summary>
+        /// Updates host viewport transform and clipping state based on current GPU state.
+        /// </summary>
+        private void UpdateViewportTransform()
+        {
+            var yControl = _state.State.YControl;
+            var face = _state.State.FaceState;
+
+            bool disableTransform = _state.State.ViewportTransformEnable == 0;
+
+            UpdateFrontFace(yControl, face.FrontFace);
+            UpdateDepthMode();
+
+            bool flipY = yControl.HasFlag(YControl.NegateY);
+
+            Span<Viewport> viewports = stackalloc Viewport[Constants.TotalViewports];
+
+            for (int index = 0; index < Constants.TotalViewports; index++)
+            {
+                if (disableTransform)
+                {
+                    ref var scissor = ref _state.State.ScreenScissorState;
+
+                    float rScale = _channel.TextureManager.RenderTargetScale;
+                    var scissorRect = new Rectangle<float>(0, 0, (scissor.X + scissor.Width) * rScale, (scissor.Y + scissor.Height) * rScale);
+
+                    viewports[index] = new Viewport(scissorRect, ViewportSwizzle.PositiveX, ViewportSwizzle.PositiveY, ViewportSwizzle.PositiveZ, ViewportSwizzle.PositiveW, 0, 1);
+                    continue;
+                }
+
+                ref var transform = ref _state.State.ViewportTransform[index];
+                ref var extents = ref _state.State.ViewportExtents[index];
+
+                float scaleX = MathF.Abs(transform.ScaleX);
+                float scaleY = transform.ScaleY;
+
+                if (flipY)
+                {
+                    scaleY = -scaleY;
+                }
+
+                if (!_context.Capabilities.SupportsViewportSwizzle && transform.UnpackSwizzleY() == ViewportSwizzle.NegativeY)
+                {
+                    scaleY = -scaleY;
+                }
+
+                float x = transform.TranslateX - scaleX;
+                float y = transform.TranslateY - scaleY;
+
+                float width = scaleX * 2;
+                float height = scaleY * 2;
+
+                float scale = _channel.TextureManager.RenderTargetScale;
+                if (scale != 1f)
+                {
+                    x *= scale;
+                    y *= scale;
+                    width *= scale;
+                    height *= scale;
+                }
+
+                Rectangle<float> region = new Rectangle<float>(x, y, width, height);
+
+                ViewportSwizzle swizzleX = transform.UnpackSwizzleX();
+                ViewportSwizzle swizzleY = transform.UnpackSwizzleY();
+                ViewportSwizzle swizzleZ = transform.UnpackSwizzleZ();
+                ViewportSwizzle swizzleW = transform.UnpackSwizzleW();
+
+                float depthNear = extents.DepthNear;
+                float depthFar = extents.DepthFar;
+
+                if (transform.ScaleZ < 0)
+                {
+                    float temp = depthNear;
+                    depthNear = depthFar;
+                    depthFar = temp;
+                }
+
+                viewports[index] = new Viewport(region, swizzleX, swizzleY, swizzleZ, swizzleW, depthNear, depthFar);
+            }
+
+            _context.Renderer.Pipeline.SetDepthMode(GetDepthMode());
+            _context.Renderer.Pipeline.SetViewports(viewports, disableTransform);
+
+            _currentSpecState.SetViewportTransformDisable(_state.State.ViewportTransformEnable == 0);
+            _currentSpecState.SetDepthMode(GetDepthMode() == DepthMode.MinusOneToOne);
+        }
+
+        /// <summary>
+        /// Updates the depth mode (0 to 1 or -1 to 1) based on the current viewport and depth mode register state.
+        /// </summary>
+        private void UpdateDepthMode()
+        {
+            _context.Renderer.Pipeline.SetDepthMode(GetDepthMode());
+        }
+
+        /// <summary>
+        /// Updates polygon mode state based on current GPU state.
+        /// </summary>
+        private void UpdatePolygonMode()
+        {
+            _context.Renderer.Pipeline.SetPolygonMode(_state.State.PolygonModeFront, _state.State.PolygonModeBack);
+        }
+
+        /// <summary>
+        /// Updates host depth bias (also called polygon offset) state based on current GPU state.
+        /// </summary>
+        private void UpdateDepthBiasState()
+        {
+            var depthBias = _state.State.DepthBiasState;
+
+            float factor = _state.State.DepthBiasFactor;
+            float units = _state.State.DepthBiasUnits;
+            float clamp = _state.State.DepthBiasClamp;
+
+            PolygonModeMask enables;
+
+            enables = (depthBias.PointEnable ? PolygonModeMask.Point : 0);
+            enables |= (depthBias.LineEnable ? PolygonModeMask.Line : 0);
+            enables |= (depthBias.FillEnable ? PolygonModeMask.Fill : 0);
+
+            _pipeline.BiasEnable = enables;
+            _context.Renderer.Pipeline.SetDepthBias(enables, factor, units / 2f, clamp);
+        }
+
+        /// <summary>
+        /// Updates host stencil test state based on current GPU state.
+        /// </summary>
+        private void UpdateStencilTestState()
+        {
+            var backMasks = _state.State.StencilBackMasks;
+            var test = _state.State.StencilTestState;
+            var backTest = _state.State.StencilBackTestState;
+
+            CompareOp backFunc;
+            StencilOp backSFail;
+            StencilOp backDpPass;
+            StencilOp backDpFail;
+            int backFuncRef;
+            int backFuncMask;
+            int backMask;
+
+            if (backTest.TwoSided)
+            {
+                backFunc = backTest.BackFunc;
+                backSFail = backTest.BackSFail;
+                backDpPass = backTest.BackDpPass;
+                backDpFail = backTest.BackDpFail;
+                backFuncRef = backMasks.FuncRef;
+                backFuncMask = backMasks.FuncMask;
+                backMask = backMasks.Mask;
+            }
+            else
+            {
+                backFunc = test.FrontFunc;
+                backSFail = test.FrontSFail;
+                backDpPass = test.FrontDpPass;
+                backDpFail = test.FrontDpFail;
+                backFuncRef = test.FrontFuncRef;
+                backFuncMask = test.FrontFuncMask;
+                backMask = test.FrontMask;
+            }
+
+            StencilTestDescriptor descriptor = new StencilTestDescriptor(
+                test.Enable,
+                test.FrontFunc,
+                test.FrontSFail,
+                test.FrontDpPass,
+                test.FrontDpFail,
+                test.FrontFuncRef,
+                test.FrontFuncMask,
+                test.FrontMask,
+                backFunc,
+                backSFail,
+                backDpPass,
+                backDpFail,
+                backFuncRef,
+                backFuncMask,
+                backMask);
+
+            _pipeline.StencilTest = descriptor;
+            _context.Renderer.Pipeline.SetStencilTest(descriptor);
+        }
+
+        /// <summary>
+        /// Updates user-defined clipping based on the guest GPU state.
+        /// </summary>
+        private void UpdateUserClipState()
+        {
+            uint clipMask = _state.State.ClipDistanceEnable & _vsClipDistancesWritten;
+
+            for (int i = 0; i < Constants.TotalClipDistances; ++i)
+            {
+                _context.Renderer.Pipeline.SetUserClipDistance(i, (clipMask & (1 << i)) != 0);
+            }
+        }
+
+        /// <summary>
+        /// Updates current sampler pool address and size based on guest GPU state.
+        /// </summary>
+        private void UpdateSamplerPoolState()
+        {
+            var texturePool = _state.State.TexturePoolState;
+            var samplerPool = _state.State.SamplerPoolState;
+
+            var samplerIndex = _state.State.SamplerIndex;
+
+            int maximumId = samplerIndex == SamplerIndex.ViaHeaderIndex
+                ? texturePool.MaximumId
+                : samplerPool.MaximumId;
+
+            _channel.TextureManager.SetGraphicsSamplerPool(samplerPool.Address.Pack(), maximumId, samplerIndex);
+        }
+
+        /// <summary>
+        /// Updates current texture pool address and size based on guest GPU state.
+        /// </summary>
+        private void UpdateTexturePoolState()
+        {
+            var texturePool = _state.State.TexturePoolState;
+
+            _channel.TextureManager.SetGraphicsTexturePool(texturePool.Address.Pack(), texturePool.MaximumId);
+            _channel.TextureManager.SetGraphicsTextureBufferIndex((int)_state.State.TextureBufferIndex);
+
+            _currentSpecState.SetPoolState(GetPoolState());
+        }
+
+        /// <summary>
+        /// Updates host vertex attributes based on guest GPU state.
+        /// </summary>
+        private void UpdateVertexAttribState()
+        {
+            uint vbEnableMask = _vbEnableMask;
+
+            Span<VertexAttribDescriptor> vertexAttribs = stackalloc VertexAttribDescriptor[Constants.TotalVertexAttribs];
+
+            for (int index = 0; index < Constants.TotalVertexAttribs; index++)
+            {
+                var vertexAttrib = _state.State.VertexAttribState[index];
+
+                int bufferIndex = vertexAttrib.UnpackBufferIndex();
+
+                if ((vbEnableMask & (1u << bufferIndex)) == 0)
+                {
+                    // Using a vertex buffer that doesn't exist is invalid, so let's use a dummy attribute for those cases.
+                    vertexAttribs[index] = new VertexAttribDescriptor(0, 0, true, Format.R32G32B32A32Float);
+                    continue;
+                }
+
+                if (!FormatTable.TryGetAttribFormat(vertexAttrib.UnpackFormat(), out Format format))
+                {
+                    Logger.Debug?.Print(LogClass.Gpu, $"Invalid attribute format 0x{vertexAttrib.UnpackFormat():X}.");
+
+                    format = vertexAttrib.UnpackType() switch
+                    {
+                        VertexAttribType.Sint => Format.R32G32B32A32Sint,
+                        VertexAttribType.Uint => Format.R32G32B32A32Uint,
+                        _ => Format.R32G32B32A32Float
+                    };
+                }
+
+                vertexAttribs[index] = new VertexAttribDescriptor(
+                    bufferIndex,
+                    vertexAttrib.UnpackOffset(),
+                    vertexAttrib.UnpackIsConstant(),
+                    format);
+            }
+
+            _pipeline.SetVertexAttribs(vertexAttribs);
+            _context.Renderer.Pipeline.SetVertexAttribs(vertexAttribs);
+            _currentSpecState.SetAttributeTypes(ref _state.State.VertexAttribState);
+        }
+
+        /// <summary>
+        /// Updates host line width based on guest GPU state.
+        /// </summary>
+        private void UpdateLineState()
+        {
+            float width = _state.State.LineWidthSmooth;
+            bool smooth = _state.State.LineSmoothEnable;
+
+            _pipeline.LineWidth = width;
+            _context.Renderer.Pipeline.SetLineParameters(width, smooth);
+        }
+
+        /// <summary>
+        /// Updates host point size based on guest GPU state.
+        /// </summary>
+        private void UpdatePointState()
+        {
+            float size = _state.State.PointSize;
+            bool isProgramPointSize = _state.State.VertexProgramPointSize;
+            bool enablePointSprite = _state.State.PointSpriteEnable;
+
+            // TODO: Need to figure out a way to map PointCoordReplace enable bit.
+            Origin origin = (_state.State.PointCoordReplace & 4) == 0 ? Origin.LowerLeft : Origin.UpperLeft;
+
+            _context.Renderer.Pipeline.SetPointParameters(size, isProgramPointSize, enablePointSprite, origin);
+
+            _currentSpecState.SetProgramPointSizeEnable(isProgramPointSize);
+            _currentSpecState.SetPointSize(size);
+        }
+
+        /// <summary>
+        /// Updates host primitive restart based on guest GPU state.
+        /// </summary>
+        private void UpdatePrimitiveRestartState()
+        {
+            PrimitiveRestartState primitiveRestart = _state.State.PrimitiveRestartState;
+            bool enable = primitiveRestart.Enable && (_drawState.DrawIndexed || _state.State.PrimitiveRestartDrawArrays);
+
+            _pipeline.PrimitiveRestartEnable = enable;
+            _context.Renderer.Pipeline.SetPrimitiveRestart(enable, primitiveRestart.Index);
+        }
+
+        /// <summary>
+        /// Updates host index buffer binding based on guest GPU state.
+        /// </summary>
+        private void UpdateIndexBufferState()
+        {
+            var indexBuffer = _state.State.IndexBufferState;
+
+            if (_drawState.IndexCount == 0)
+            {
+                return;
+            }
+
+            ulong gpuVa = indexBuffer.Address.Pack();
+
+            // Do not use the end address to calculate the size, because
+            // the result may be much larger than the real size of the index buffer.
+            ulong size = (ulong)(_drawState.FirstIndex + _drawState.IndexCount);
+
+            switch (indexBuffer.Type)
+            {
+                case IndexType.UShort: size *= 2; break;
+                case IndexType.UInt: size *= 4; break;
+            }
+
+            _channel.BufferManager.SetIndexBuffer(gpuVa, size, indexBuffer.Type);
+        }
+
+        /// <summary>
+        /// Updates host vertex buffer bindings based on guest GPU state.
+        /// </summary>
+        private void UpdateVertexBufferState()
+        {
+            IndexType indexType = _state.State.IndexBufferState.Type;
+            bool indexTypeSmall = indexType == IndexType.UByte || indexType == IndexType.UShort;
+
+            _drawState.IsAnyVbInstanced = false;
+
+            bool drawIndexed = _drawState.DrawIndexed;
+            bool drawIndirect = _drawState.DrawIndirect;
+            int drawFirstVertex = _drawState.DrawFirstVertex;
+            int drawVertexCount = _drawState.DrawVertexCount;
+            uint vbEnableMask = 0;
+
+            for (int index = 0; index < Constants.TotalVertexBuffers; index++)
+            {
+                var vertexBuffer = _state.State.VertexBufferState[index];
+
+                if (!vertexBuffer.UnpackEnable())
+                {
+                    _pipeline.VertexBuffers[index] = new BufferPipelineDescriptor(false, 0, 0);
+                    _channel.BufferManager.SetVertexBuffer(index, 0, 0, 0, 0);
+
+                    continue;
+                }
+
+                GpuVa endAddress = _state.State.VertexBufferEndAddress[index];
+
+                ulong address = vertexBuffer.Address.Pack();
+
+                if (_channel.MemoryManager.IsMapped(address))
+                {
+                    vbEnableMask |= 1u << index;
+                }
+
+                int stride = vertexBuffer.UnpackStride();
+
+                bool instanced = _state.State.VertexBufferInstanced[index];
+
+                int divisor = instanced ? vertexBuffer.Divisor : 0;
+
+                _drawState.IsAnyVbInstanced |= divisor != 0;
+
+                ulong vbSize = endAddress.Pack() - address + 1;
+                ulong size;
+
+                if (_drawState.IbStreamer.HasInlineIndexData || drawIndexed || stride == 0 || instanced)
+                {
+                    // This size may be (much) larger than the real vertex buffer size.
+                    // Avoid calculating it this way, unless we don't have any other option.
+
+                    size = vbSize;
+
+                    if (stride > 0 && indexTypeSmall && drawIndexed && !drawIndirect && !instanced)
+                    {
+                        // If the index type is a small integer type, then we might be still able
+                        // to reduce the vertex buffer size based on the maximum possible index value.
+
+                        ulong maxVertexBufferSize = indexType == IndexType.UByte ? 0x100UL : 0x10000UL;
+
+                        maxVertexBufferSize += _state.State.FirstVertex;
+                        maxVertexBufferSize *= (uint)stride;
+
+                        size = Math.Min(size, maxVertexBufferSize);
+                    }
+                }
+                else
+                {
+                    // For non-indexed draws, we can guess the size from the vertex count
+                    // and stride.
+
+                    int firstInstance = (int)_state.State.FirstInstance;
+
+                    size = Math.Min(vbSize, (ulong)((firstInstance + drawFirstVertex + drawVertexCount) * stride));
+                }
+
+                _pipeline.VertexBuffers[index] = new BufferPipelineDescriptor(_channel.MemoryManager.IsMapped(address), stride, divisor);
+                _channel.BufferManager.SetVertexBuffer(index, address, size, stride, divisor);
+            }
+
+            if (_vbEnableMask != vbEnableMask)
+            {
+                _vbEnableMask = vbEnableMask;
+                UpdateVertexAttribState();
+            }
+        }
+
+        /// <summary>
+        /// Updates host face culling and orientation based on guest GPU state.
+        /// </summary>
+        private void UpdateFaceState()
+        {
+            var yControl = _state.State.YControl;
+            var face = _state.State.FaceState;
+
+            _pipeline.CullEnable = face.CullEnable;
+            _pipeline.CullMode = face.CullFace;
+            _context.Renderer.Pipeline.SetFaceCulling(face.CullEnable, face.CullFace);
+
+            UpdateFrontFace(yControl, face.FrontFace);
+        }
+
+        /// <summary>
+        /// Updates the front face based on the current front face and the origin.
+        /// </summary>
+        /// <param name="yControl">Y control register value, where the origin is located</param>
+        /// <param name="frontFace">Front face</param>
+        private void UpdateFrontFace(YControl yControl, FrontFace frontFace)
+        {
+            bool isUpperLeftOrigin = !yControl.HasFlag(YControl.TriangleRastFlip);
+
+            if (isUpperLeftOrigin)
+            {
+                frontFace = frontFace == FrontFace.CounterClockwise ? FrontFace.Clockwise : FrontFace.CounterClockwise;
+            }
+
+            _pipeline.FrontFace = frontFace;
+            _context.Renderer.Pipeline.SetFrontFace(frontFace);
+        }
+
+        /// <summary>
+        /// Updates host render target color masks, based on guest GPU state.
+        /// This defines which color channels are written to each color buffer.
+        /// </summary>
+        private void UpdateRtColorMask()
+        {
+            bool rtColorMaskShared = _state.State.RtColorMaskShared;
+
+            Span<uint> componentMasks = stackalloc uint[Constants.TotalRenderTargets];
+
+            for (int index = 0; index < Constants.TotalRenderTargets; index++)
+            {
+                var colorMask = _state.State.RtColorMask[rtColorMaskShared ? 0 : index];
+
+                uint componentMask;
+
+                componentMask = (colorMask.UnpackRed() ? 1u : 0u);
+                componentMask |= (colorMask.UnpackGreen() ? 2u : 0u);
+                componentMask |= (colorMask.UnpackBlue() ? 4u : 0u);
+                componentMask |= (colorMask.UnpackAlpha() ? 8u : 0u);
+
+                componentMasks[index] = componentMask;
+                _pipeline.ColorWriteMask[index] = componentMask;
+            }
+
+            _context.Renderer.Pipeline.SetRenderTargetColorMasks(componentMasks);
+        }
+
+        /// <summary>
+        /// Updates host render target color buffer blending state, based on guest state.
+        /// </summary>
+        private void UpdateBlendState()
+        {
+            if (_state.State.BlendUcodeEnable != BlendUcodeEnable.Disabled)
+            {
+                if (_context.Capabilities.SupportsBlendEquationAdvanced && _blendManager.TryGetAdvancedBlend(out var blendDescriptor))
+                {
+                    // Try to HLE it using advanced blend on the host if we can.
+                    _context.Renderer.Pipeline.SetBlendState(blendDescriptor);
+                    return;
+                }
+                else
+                {
+                    // TODO: Blend emulation fallback.
+                }
+            }
+
+            bool blendIndependent = _state.State.BlendIndependent;
+            ColorF blendConstant = _state.State.BlendConstant;
+
+            bool dualSourceBlendEnabled = false;
+
+            if (blendIndependent)
+            {
+                for (int index = 0; index < Constants.TotalRenderTargets; index++)
+                {
+                    bool enable = _state.State.BlendEnable[index];
+                    var blend = _state.State.BlendState[index];
+
+                    var descriptor = new BlendDescriptor(
+                        enable,
+                        blendConstant,
+                        blend.ColorOp,
+                        FilterBlendFactor(blend.ColorSrcFactor, index),
+                        FilterBlendFactor(blend.ColorDstFactor, index),
+                        blend.AlphaOp,
+                        FilterBlendFactor(blend.AlphaSrcFactor, index),
+                        FilterBlendFactor(blend.AlphaDstFactor, index));
+
+                    if (enable &&
+                        (blend.ColorSrcFactor.IsDualSource() ||
+                        blend.ColorDstFactor.IsDualSource() ||
+                        blend.AlphaSrcFactor.IsDualSource() ||
+                        blend.AlphaDstFactor.IsDualSource()))
+                    {
+                        dualSourceBlendEnabled = true;
+                    }
+
+                    _pipeline.BlendDescriptors[index] = descriptor;
+                    _context.Renderer.Pipeline.SetBlendState(index, descriptor);
+                }
+            }
+            else
+            {
+                bool enable = _state.State.BlendEnable[0];
+                var blend = _state.State.BlendStateCommon;
+
+                var descriptor = new BlendDescriptor(
+                    enable,
+                    blendConstant,
+                    blend.ColorOp,
+                    FilterBlendFactor(blend.ColorSrcFactor, 0),
+                    FilterBlendFactor(blend.ColorDstFactor, 0),
+                    blend.AlphaOp,
+                    FilterBlendFactor(blend.AlphaSrcFactor, 0),
+                    FilterBlendFactor(blend.AlphaDstFactor, 0));
+
+                if (enable &&
+                    (blend.ColorSrcFactor.IsDualSource() ||
+                    blend.ColorDstFactor.IsDualSource() ||
+                    blend.AlphaSrcFactor.IsDualSource() ||
+                    blend.AlphaDstFactor.IsDualSource()))
+                {
+                    dualSourceBlendEnabled = true;
+                }
+
+                for (int index = 0; index < Constants.TotalRenderTargets; index++)
+                {
+                    _pipeline.BlendDescriptors[index] = descriptor;
+                    _context.Renderer.Pipeline.SetBlendState(index, descriptor);
+                }
+            }
+
+            _currentSpecState.SetDualSourceBlendEnabled(dualSourceBlendEnabled);
+        }
+
+        /// <summary>
+        /// Gets a blend factor for the color target currently.
+        /// This will return <paramref name="factor"/> unless the target format has no alpha component,
+        /// in which case it will replace destination alpha factor with a constant factor of one or zero.
+        /// </summary>
+        /// <param name="factor">Input factor</param>
+        /// <param name="index">Color target index</param>
+        /// <returns>New blend factor</returns>
+        private BlendFactor FilterBlendFactor(BlendFactor factor, int index)
+        {
+            // If any color target format without alpha is being used, we need to make sure that
+            // if blend is active, it will not use destination alpha as a factor.
+            // That is required because RGBX formats are emulated using host RGBA formats.
+
+            if (_state.State.RtColorState[index].Format.NoAlpha())
+            {
+                switch (factor)
+                {
+                    case BlendFactor.DstAlpha:
+                    case BlendFactor.DstAlphaGl:
+                        factor = BlendFactor.One;
+                        break;
+                    case BlendFactor.OneMinusDstAlpha:
+                    case BlendFactor.OneMinusDstAlphaGl:
+                        factor = BlendFactor.Zero;
+                        break;
+                }
+            }
+
+            return factor;
+        }
+
+        /// <summary>
+        /// Updates host logical operation state, based on guest state.
+        /// </summary>
+        private void UpdateLogicOpState()
+        {
+            LogicalOpState logicOpState = _state.State.LogicOpState;
+
+            _pipeline.SetLogicOpState(logicOpState.Enable, logicOpState.LogicalOp);
+            _context.Renderer.Pipeline.SetLogicOpState(logicOpState.Enable, logicOpState.LogicalOp);
+        }
+
+        /// <summary>
+        /// Updates multisample state, based on guest state.
+        /// </summary>
+        private void UpdateMultisampleState()
+        {
+            bool alphaToCoverageEnable = (_state.State.MultisampleControl & 1) != 0;
+            bool alphaToOneEnable = (_state.State.MultisampleControl & 0x10) != 0;
+
+            _context.Renderer.Pipeline.SetMultisampleState(new MultisampleDescriptor(
+                alphaToCoverageEnable,
+                _state.State.AlphaToCoverageDitherEnable,
+                alphaToOneEnable));
+
+            _currentSpecState.SetAlphaToCoverageEnable(alphaToCoverageEnable, _state.State.AlphaToCoverageDitherEnable);
+        }
+
+        /// <summary>
+        /// Updates the early z flag, based on guest state.
+        /// </summary>
+        private void UpdateEarlyZState()
+        {
+            _currentSpecState.SetEarlyZForce(_state.State.EarlyZForce);
+        }
+
+        /// <summary>
+        /// Updates host shaders based on the guest GPU state.
+        /// </summary>
+        private void UpdateShaderState()
+        {
+            var shaderCache = _channel.MemoryManager.Physical.ShaderCache;
+
+            _vtgWritesRtLayer = false;
+
+            ShaderAddresses addresses = new ShaderAddresses();
+            Span<ulong> addressesSpan = addresses.AsSpan();
+
+            ulong baseAddress = _state.State.ShaderBaseAddress.Pack();
+
+            for (int index = 0; index < 6; index++)
+            {
+                var shader = _state.State.ShaderState[index];
+                if (!shader.UnpackEnable() && index != 1)
+                {
+                    continue;
+                }
+
+                addressesSpan[index] = baseAddress + shader.Offset;
+            }
+
+            CachedShaderProgram gs = shaderCache.GetGraphicsShader(ref _state.State, ref _pipeline, _channel, ref _currentSpecState.GetPoolState(), ref _currentSpecState.GetGraphicsState(), addresses);
+
+            // Consume the modified flag for spec state so that it isn't checked again.
+            _currentSpecState.SetShader(gs);
+
+            _shaderSpecState = gs.SpecializationState;
+
+            byte oldVsClipDistancesWritten = _vsClipDistancesWritten;
+
+            _drawState.VsUsesInstanceId = gs.Shaders[1]?.Info.UsesInstanceId ?? false;
+            _vsUsesDrawParameters = gs.Shaders[1]?.Info.UsesDrawParameters ?? false;
+            _vsClipDistancesWritten = gs.Shaders[1]?.Info.ClipDistancesWritten ?? 0;
+
+            if (oldVsClipDistancesWritten != _vsClipDistancesWritten)
+            {
+                UpdateUserClipState();
+            }
+
+            UpdateShaderBindings(gs.Bindings);
+
+            for (int stageIndex = 0; stageIndex < Constants.ShaderStages; stageIndex++)
+            {
+                ShaderProgramInfo info = gs.Shaders[stageIndex + 1]?.Info;
+
+                if (info?.UsesRtLayer == true)
+                {
+                    _vtgWritesRtLayer = true;
+                }
+
+                _currentProgramInfo[stageIndex] = info;
+            }
+
+            _context.Renderer.Pipeline.SetProgram(gs.HostProgram);
+        }
+
+        /// <summary>
+        /// Updates bindings consumed by the shader on the texture and buffer managers.
+        /// </summary>
+        /// <param name="bindings">Bindings for the active shader</param>
+        private void UpdateShaderBindings(CachedShaderBindings bindings)
+        {
+            _channel.TextureManager.SetGraphicsBindings(bindings);
+            _channel.BufferManager.SetGraphicsBufferBindings(bindings);
+        }
+
+        /// <summary>
+        /// Gets the current texture pool state.
+        /// </summary>
+        /// <returns>Texture pool state</returns>
+        private GpuChannelPoolState GetPoolState()
+        {
+            return new GpuChannelPoolState(
+                _state.State.TexturePoolState.Address.Pack(),
+                _state.State.TexturePoolState.MaximumId,
+                (int)_state.State.TextureBufferIndex);
+        }
+
+        /// <summary>
+        /// Gets the depth mode that is currently being used (zero to one or minus one to one).
+        /// </summary>
+        /// <returns>Current depth mode</returns>
+        private DepthMode GetDepthMode()
+        {
+            ref var transform = ref _state.State.ViewportTransform[0];
+            ref var extents = ref _state.State.ViewportExtents[0];
+
+            DepthMode depthMode;
+
+            if (!float.IsInfinity(extents.DepthNear) &&
+                !float.IsInfinity(extents.DepthFar) &&
+                (extents.DepthFar - extents.DepthNear) != 0)
+            {
+                // Try to guess the depth mode being used on the high level API
+                // based on current transform.
+                // It is setup like so by said APIs:
+                // If depth mode is ZeroToOne:
+                //  TranslateZ = Near
+                //  ScaleZ = Far - Near
+                // If depth mode is MinusOneToOne:
+                //  TranslateZ = (Near + Far) / 2
+                //  ScaleZ = (Far - Near) / 2
+                // DepthNear/Far are sorted such as that Near is always less than Far.
+                depthMode = extents.DepthNear != transform.TranslateZ &&
+                            extents.DepthFar  != transform.TranslateZ
+                    ? DepthMode.MinusOneToOne
+                    : DepthMode.ZeroToOne;
+            }
+            else
+            {
+                // If we can't guess from the viewport transform, then just use the depth mode register.
+                depthMode = (DepthMode)(_state.State.DepthMode & 1);
+            }
+
+            return depthMode;
+        }
+
+        /// <summary>
+        /// Forces the shaders to be rebound on the next draw.
+        /// </summary>
+        public void ForceShaderUpdate()
+        {
+            _updateTracker.ForceDirty(ShaderStateIndex);
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClass.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClass.cs
new file mode 100644
index 00000000..caeee18e
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClass.cs
@@ -0,0 +1,620 @@
+using Ryujinx.Graphics.Device;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.GPFifo;
+using Ryujinx.Graphics.Gpu.Engine.InlineToMemory;
+using Ryujinx.Graphics.Gpu.Engine.Threed.Blender;
+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 GPFifoClass _fifoClass;
+        private readonly DeviceStateWithShadow<ThreedClassState> _state;
+
+        private readonly InlineToMemoryClass _i2mClass;
+        private readonly AdvancedBlendManager _blendManager;
+        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, GPFifoClass fifoClass)
+        {
+            _context = context;
+            _fifoClass = fifoClass;
+            _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.InvalidateSamplerCacheNoWfi), new RwCallback(InvalidateSamplerCacheNoWfi, null) },
+                { nameof(ThreedClassState.InvalidateTextureHeaderCacheNoWfi), new RwCallback(InvalidateTextureHeaderCacheNoWfi, null) },
+                { nameof(ThreedClassState.TextureBarrier), new RwCallback(TextureBarrier, null) },
+                { nameof(ThreedClassState.LoadBlendUcodeStart), new RwCallback(LoadBlendUcodeStart, null) },
+                { nameof(ThreedClassState.LoadBlendUcodeInstruction), new RwCallback(LoadBlendUcodeInstruction, null) },
+                { nameof(ThreedClassState.TextureBarrierTiled), new RwCallback(TextureBarrierTiled, null) },
+                { nameof(ThreedClassState.DrawTextureSrcY), new RwCallback(DrawTexture, null) },
+                { nameof(ThreedClassState.DrawVertexArrayBeginEndInstanceFirst), new RwCallback(DrawVertexArrayBeginEndInstanceFirst, null) },
+                { nameof(ThreedClassState.DrawVertexArrayBeginEndInstanceSubsequent), new RwCallback(DrawVertexArrayBeginEndInstanceSubsequent, 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.DrawIndexBuffer32BeginEndInstanceFirst), new RwCallback(DrawIndexBuffer32BeginEndInstanceFirst, null) },
+                { nameof(ThreedClassState.DrawIndexBuffer16BeginEndInstanceFirst), new RwCallback(DrawIndexBuffer16BeginEndInstanceFirst, null) },
+                { nameof(ThreedClassState.DrawIndexBuffer8BeginEndInstanceFirst), new RwCallback(DrawIndexBuffer8BeginEndInstanceFirst, null) },
+                { nameof(ThreedClassState.DrawIndexBuffer32BeginEndInstanceSubsequent), new RwCallback(DrawIndexBuffer32BeginEndInstanceSubsequent, null) },
+                { nameof(ThreedClassState.DrawIndexBuffer16BeginEndInstanceSubsequent), new RwCallback(DrawIndexBuffer16BeginEndInstanceSubsequent, null) },
+                { nameof(ThreedClassState.DrawIndexBuffer8BeginEndInstanceSubsequent), new RwCallback(DrawIndexBuffer8BeginEndInstanceSubsequent, 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 spec = new SpecializationStateUpdater(context);
+            var drawState = new DrawState();
+
+            _drawManager = new DrawManager(context, channel, _state, drawState, spec);
+            _blendManager = new AdvancedBlendManager(_state);
+            _semaphoreUpdater = new SemaphoreUpdater(context, channel, _state);
+            _cbUpdater = new ConstantBufferUpdater(channel, _state);
+            _stateUpdater = new StateUpdater(context, channel, _state, drawState, _blendManager, spec);
+
+            // 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()
+        {
+            _fifoClass.CreatePendingSyncs();
+            _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="updateFlags">Flags indicating which render targets should be updated and how</param>
+        /// <param name="singleUse">If this is not -1, it indicates that only the given indexed target will be used.</param>
+        public void UpdateRenderTargetState(RenderTargetUpdateFlags updateFlags, int singleUse = -1)
+        {
+            _stateUpdater.UpdateRenderTargetState(updateFlags, singleUse);
+        }
+
+        /// <summary>
+        /// Updates scissor based on current render target state.
+        /// </summary>
+        public void UpdateScissorState()
+        {
+            _stateUpdater.UpdateScissorState();
+        }
+
+        /// <summary>
+        /// Marks the entire state as dirty, forcing a full host state update before the next draw.
+        /// </summary>
+        public void ForceStateDirty()
+        {
+            _drawManager.ForceStateDirty();
+            _stateUpdater.SetAllDirty();
+        }
+
+        /// <summary>
+        /// Marks the specified register offset as dirty, forcing the associated state to update on the next draw.
+        /// </summary>
+        /// <param name="offset">Register offset</param>
+        public void ForceStateDirty(int offset)
+        {
+            _stateUpdater.SetDirty(offset);
+        }
+
+        /// <summary>
+        /// Forces the shaders to be rebound on the next draw.
+        /// </summary>
+        public void ForceShaderUpdate()
+        {
+            _stateUpdater.ForceShaderUpdate();
+        }
+
+        /// <summary>
+        /// Create any syncs from WaitForIdle command that are currently pending.
+        /// </summary>
+        public void CreatePendingSyncs()
+        {
+            _fifoClass.CreatePendingSyncs();
+        }
+
+        /// <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 block of data to the Inline-to-Memory engine.
+        /// </summary>
+        /// <param name="data">Data to push</param>
+        public void LoadInlineData(ReadOnlySpan<int> data)
+        {
+            _i2mClass.LoadInlineData(data);
+        }
+
+        /// <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.AdvanceSequence();
+            _context.CreateHostSyncIfNeeded(true, true);
+            _context.Renderer.UpdateCounters(); // Poll the query counters, the game may want an updated result.
+            _context.Synchronization.IncrementSyncpoint(syncpointId);
+        }
+
+        /// <summary>
+        /// Invalidates the cache with the sampler descriptors from the sampler pool.
+        /// </summary>
+        /// <param name="argument">Method call argument (unused)</param>
+        private void InvalidateSamplerCacheNoWfi(int argument)
+        {
+            _context.AdvanceSequence();
+        }
+
+        /// <summary>
+        /// Invalidates the cache with the texture descriptors from the texture pool.
+        /// </summary>
+        /// <param name="argument">Method call argument (unused)</param>
+        private void InvalidateTextureHeaderCacheNoWfi(int argument)
+        {
+            _context.AdvanceSequence();
+        }
+
+        /// <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>
+        /// Sets the start offset of the blend microcode in memory.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void LoadBlendUcodeStart(int argument)
+        {
+            _blendManager.LoadBlendUcodeStart(argument);
+        }
+
+        /// <summary>
+        /// Pushes one word of blend microcode.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void LoadBlendUcodeInstruction(int argument)
+        {
+            _blendManager.LoadBlendUcodeInstruction(argument);
+        }
+
+        /// <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>
+        /// Draws a texture, without needing to specify shader programs.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawTexture(int argument)
+        {
+            _drawManager.DrawTexture(this, argument);
+        }
+
+        /// <summary>
+        /// Performs a non-indexed draw with the specified topology, index and count.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawVertexArrayBeginEndInstanceFirst(int argument)
+        {
+            _drawManager.DrawVertexArrayBeginEndInstanceFirst(this, argument);
+        }
+
+        /// <summary>
+        /// Performs a non-indexed draw with the specified topology, index and count,
+        /// while incrementing the current instance.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawVertexArrayBeginEndInstanceSubsequent(int argument)
+        {
+            _drawManager.DrawVertexArrayBeginEndInstanceSubsequent(this, argument);
+        }
+
+        /// <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 8-bit index buffer elements.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawIndexBuffer8BeginEndInstanceFirst(int argument)
+        {
+            _drawManager.DrawIndexBuffer8BeginEndInstanceFirst(this, argument);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with 16-bit index buffer elements.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawIndexBuffer16BeginEndInstanceFirst(int argument)
+        {
+            _drawManager.DrawIndexBuffer16BeginEndInstanceFirst(this, argument);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with 32-bit index buffer elements.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawIndexBuffer32BeginEndInstanceFirst(int argument)
+        {
+            _drawManager.DrawIndexBuffer32BeginEndInstanceFirst(this, argument);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with 8-bit index buffer elements,
+        /// while also pre-incrementing the current instance value.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawIndexBuffer8BeginEndInstanceSubsequent(int argument)
+        {
+            _drawManager.DrawIndexBuffer8BeginEndInstanceSubsequent(this, argument);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with 16-bit index buffer elements,
+        /// while also pre-incrementing the current instance value.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawIndexBuffer16BeginEndInstanceSubsequent(int argument)
+        {
+            _drawManager.DrawIndexBuffer16BeginEndInstanceSubsequent(this, argument);
+        }
+
+        /// <summary>
+        /// Performs a indexed draw with 32-bit index buffer elements,
+        /// while also pre-incrementing the current instance value.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void DrawIndexBuffer32BeginEndInstanceSubsequent(int argument)
+        {
+            _drawManager.DrawIndexBuffer32BeginEndInstanceSubsequent(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;
+        }
+
+        /// <summary>
+        /// Performs a indexed or non-indexed draw.
+        /// </summary>
+        /// <param name="topology">Primitive topology</param>
+        /// <param name="count">Index count for indexed draws, vertex count for non-indexed draws</param>
+        /// <param name="instanceCount">Instance count</param>
+        /// <param name="firstIndex">First index on the index buffer for indexed draws, ignored for non-indexed draws</param>
+        /// <param name="firstVertex">First vertex on the vertex buffer</param>
+        /// <param name="firstInstance">First instance</param>
+        /// <param name="indexed">True if the draw is indexed, false otherwise</param>
+        public void Draw(
+            PrimitiveTopology topology,
+            int count,
+            int instanceCount,
+            int firstIndex,
+            int firstVertex,
+            int firstInstance,
+            bool indexed)
+        {
+            _drawManager.Draw(this, topology, count, instanceCount, firstIndex, firstVertex, firstInstance, indexed);
+        }
+
+        /// <summary>
+        /// Performs a indirect draw, with parameters from a GPU buffer.
+        /// </summary>
+        /// <param name="topology">Primitive topology</param>
+        /// <param name="indirectBufferAddress">Address of the buffer with the draw parameters, such as count, first index, etc</param>
+        /// <param name="parameterBufferAddress">Address of the buffer with the draw count</param>
+        /// <param name="maxDrawCount">Maximum number of draws that can be made</param>
+        /// <param name="stride">Distance in bytes between each entry on the data pointed to by <paramref name="indirectBufferAddress"/></param>
+        /// <param name="indexCount">Maximum number of indices that the draw can consume</param>
+        /// <param name="drawType">Type of the indirect draw, which can be indexed or non-indexed, with or without a draw count</param>
+        public void DrawIndirect(
+            PrimitiveTopology topology,
+            ulong indirectBufferAddress,
+            ulong parameterBufferAddress,
+            int maxDrawCount,
+            int stride,
+            int indexCount,
+            IndirectDrawType drawType)
+        {
+            _drawManager.DrawIndirect(this, topology, indirectBufferAddress, parameterBufferAddress, maxDrawCount, stride, indexCount, drawType);
+        }
+
+        /// <summary>
+        /// Clears the current color and depth-stencil buffers.
+        /// Which buffers should be cleared can also specified with the arguments.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        /// <param name="layerCount">For array and 3D textures, indicates how many layers should be cleared</param>
+        public void Clear(int argument, int layerCount)
+        {
+            _drawManager.Clear(this, argument, layerCount);
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClassState.cs b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClassState.cs
new file mode 100644
index 00000000..8f26f38f
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Threed/ThreedClassState.cs
@@ -0,0 +1,1048 @@
+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 Ryujinx.Graphics.Shader;
+using System;
+using System.Runtime.CompilerServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Threed
+{
+    /// <summary>
+    /// Shader stage name.
+    /// </summary>
+    enum ShaderType
+    {
+        Vertex,
+        TessellationControl,
+        TessellationEvaluation,
+        Geometry,
+        Fragment
+    }
+
+    /// <summary>
+    /// Tessellation mode.
+    /// </summary>
+    struct TessMode
+    {
+#pragma warning disable CS0649
+        public uint Packed;
+#pragma warning restore CS0649
+
+        /// <summary>
+        /// Unpacks the tessellation abstract patch type.
+        /// </summary>
+        /// <returns>Abtract patch type</returns>
+        public TessPatchType UnpackPatchType()
+        {
+            return (TessPatchType)(Packed & 3);
+        }
+
+        /// <summary>
+        /// Unpacks the spacing between tessellated vertices of the patch.
+        /// </summary>
+        /// <returns>Spacing between tessellated vertices</returns>
+        public TessSpacing UnpackSpacing()
+        {
+            return (TessSpacing)((Packed >> 4) & 3);
+        }
+
+        /// <summary>
+        /// Unpacks the primitive winding order.
+        /// </summary>
+        /// <returns>True if clockwise, false if counter-clockwise</returns>
+        public bool UnpackCw()
+        {
+            return (Packed & (1 << 8)) != 0;
+        }
+    }
+
+    /// <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>
+    /// Indicates whenever the blend microcode processes RGB and alpha components.
+    /// </summary>
+    enum BlendUcodeEnable
+    {
+        Disabled = 0,
+        EnableRGB = 1,
+        EnableAlpha = 2,
+        EnableRGBA = 3
+    }
+
+    /// <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 attribute vector and component size.
+    /// </summary>
+    enum VertexAttribSize
+    {
+        Size32x4 = 1,
+        Size32x3 = 2,
+        Size16x4 = 3,
+        Size32x2 = 4,
+        Size16x3 = 5,
+        Size8x4 = 0xa,
+        Size16x2 = 0xf,
+        Size32 = 0x12,
+        Size8x3 = 0x13,
+        Size8x2 = 0x18,
+        Size16 = 0x1b,
+        Size8 = 0x1d,
+        Rgb10A2 = 0x30,
+        Rg11B10 = 0x31
+    }
+
+    /// <summary>
+    /// Vertex attribute component type.
+    /// </summary>
+    enum VertexAttribType
+    {
+        Snorm = 1,
+        Unorm = 2,
+        Sint = 3,
+        Uint = 4,
+        Uscaled = 5,
+        Sscaled = 6,
+        Float = 7
+    }
+
+    /// <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>
+        /// Unpacks the Maxwell attribute size.
+        /// </summary>
+        /// <returns>Attribute size</returns>
+        public VertexAttribSize UnpackSize()
+        {
+            return (VertexAttribSize)((Attribute >> 21) & 0x3f);
+        }
+
+        /// <summary>
+        /// Unpacks the Maxwell attribute component type.
+        /// </summary>
+        /// <returns>Attribute component type</returns>
+        public VertexAttribType UnpackType()
+        {
+            return (VertexAttribType)((Attribute >> 27) & 7);
+        }
+    }
+
+    /// <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>
+    /// RGB color components packed as 16-bit float values.
+    /// </summary>
+    struct RgbHalf
+    {
+#pragma warning disable CS0649
+        public uint R;
+        public uint G;
+        public uint B;
+        public uint Padding;
+#pragma warning restore CS0649
+
+        /// <summary>
+        /// Unpacks the red color component as a 16-bit float value.
+        /// </summary>
+        /// <returns>The component value</returns>
+        public Half UnpackR()
+        {
+            ushort value = (ushort)R;
+            return Unsafe.As<ushort, Half>(ref value);
+        }
+
+        /// <summary>
+        /// Unpacks the green color component as a 16-bit float value.
+        /// </summary>
+        /// <returns>The component value</returns>
+        public Half UnpackG()
+        {
+            ushort value = (ushort)G;
+            return Unsafe.As<ushort, Half>(ref value);
+        }
+
+        /// <summary>
+        /// Unpacks the blue color component as a 16-bit float value.
+        /// </summary>
+        /// <returns>The component value</returns>
+        public Half UnpackB()
+        {
+            ushort value = (ushort)B;
+            return Unsafe.As<ushort, Half>(ref value);
+        }
+    }
+
+    /// <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[10];
+        public uint BlendUcodeNormalizedDst;
+        public fixed uint Reserved2F8[10];
+        public TessMode TessMode;
+        public Array4<float> TessOuterLevel;
+        public Array2<float> TessInnerLevel;
+        public fixed uint Reserved33C[16];
+        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[2];
+        public PolygonMode PolygonModeFront;
+        public PolygonMode PolygonModeBack;
+        public Boolean32 PolygonSmoothEnable;
+        public fixed uint ReservedDB8[2];
+        public DepthBiasState DepthBiasState;
+        public int PatchVertices;
+        public BlendUcodeEnable BlendUcodeEnable;
+        public uint BlendUcodeSize;
+        public fixed uint ReservedDD8[2];
+        public uint TextureBarrier;
+        public uint WatchdogTimer;
+        public Boolean32 PrimitiveRestartDrawArrays;
+        public uint ReservedDEC;
+        public uint LoadBlendUcodeStart;
+        public uint LoadBlendUcodeInstruction;
+        public fixed uint ReservedDF8[2];
+        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[33];
+        public int DrawTextureDstX;
+        public int DrawTextureDstY;
+        public int DrawTextureDstWidth;
+        public int DrawTextureDstHeight;
+        public long DrawTextureDuDx;
+        public long DrawTextureDvDy;
+        public int DrawTextureSamplerId;
+        public int DrawTextureTextureId;
+        public int DrawTextureSrcX;
+        public int DrawTextureSrcY;
+        public fixed uint Reserved10B0[18];
+        public uint ClearFlags;
+        public fixed uint Reserved10FC[25];
+        public Array32<VertexAttribState> VertexAttribState;
+        public fixed uint Reserved11E0[13];
+        public uint DrawVertexArrayBeginEndInstanceFirst;
+        public uint DrawVertexArrayBeginEndInstanceSubsequent;
+        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[4];
+        public Boolean32 AlphaToCoverageDitherEnable;
+        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[27];
+        public uint InvalidateSamplerCacheNoWfi;
+        public uint InvalidateTextureHeaderCacheNoWfi;
+        public fixed uint Reserved142C[2];
+        public uint FirstVertex;
+        public uint FirstInstance;
+        public fixed uint Reserved143C[17];
+        public Array8<RgbHalf> BlendUcodeConstants;
+        public fixed uint Reserved1500[4];
+        public uint ClipDistanceEnable;
+        public uint Reserved1514;
+        public float PointSize;
+        public uint Reserved151C;
+        public Boolean32 PointSpriteEnable;
+        public fixed uint Reserved1524[3];
+        public uint ResetCounter;
+        public Boolean32 MultisampleEnable;
+        public Boolean32 RtDepthStencilEnable;
+        public uint MultisampleControl;
+        public fixed uint Reserved1540[4];
+        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 DrawIndexBuffer32BeginEndInstanceFirst;
+        public uint DrawIndexBuffer16BeginEndInstanceFirst;
+        public uint DrawIndexBuffer8BeginEndInstanceFirst;
+        public uint DrawIndexBuffer32BeginEndInstanceSubsequent;
+        public uint DrawIndexBuffer16BeginEndInstanceSubsequent;
+        public uint DrawIndexBuffer8BeginEndInstanceSubsequent;
+        public fixed uint Reserved17FC[32];
+        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
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Twod/TwodClass.cs b/src/Ryujinx.Graphics.Gpu/Engine/Twod/TwodClass.cs
new file mode 100644
index 00000000..4ce53e78
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Twod/TwodClass.cs
@@ -0,0 +1,379 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.Device;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Engine.Types;
+using Ryujinx.Graphics.Gpu.Image;
+using Ryujinx.Graphics.Texture;
+using Ryujinx.Memory;
+using System;
+using System.Collections.Generic;
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+using System.Runtime.Intrinsics;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Twod
+{
+    /// <summary>
+    /// Represents a 2D engine class.
+    /// </summary>
+    class TwodClass : IDeviceState
+    {
+        private readonly GpuChannel _channel;
+        private readonly DeviceState<TwodClassState> _state;
+
+        /// <summary>
+        /// Creates a new instance of the 2D engine class.
+        /// </summary>
+        /// <param name="channel">The channel that will make use of the engine</param>
+        public TwodClass(GpuChannel channel)
+        {
+            _channel = channel;
+            _state = new DeviceState<TwodClassState>(new Dictionary<string, RwCallback>
+            {
+                { nameof(TwodClassState.PixelsFromMemorySrcY0Int), new RwCallback(PixelsFromMemorySrcY0Int, null) }
+            });
+        }
+
+        /// <summary>
+        /// Reads data from the class registers.
+        /// </summary>
+        /// <param name="offset">Register byte offset</param>
+        /// <returns>Data at the specified offset</returns>
+        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>
+        public void Write(int offset, int data) => _state.Write(offset, data);
+
+        /// <summary>
+        /// Determines if data is compatible between the source and destination texture.
+        /// The two textures must have the same size, layout, and bytes per pixel.
+        /// </summary>
+        /// <param name="lhs">Info for the first texture</param>
+        /// <param name="rhs">Info for the second texture</param>
+        /// <param name="lhsFormat">Format of the first texture</param>
+        /// <param name="rhsFormat">Format of the second texture</param>
+        /// <returns>True if the data is compatible, false otherwise</returns>
+        private bool IsDataCompatible(TwodTexture lhs, TwodTexture rhs, FormatInfo lhsFormat, FormatInfo rhsFormat)
+        {
+            if (lhsFormat.BytesPerPixel != rhsFormat.BytesPerPixel ||
+                lhs.Height != rhs.Height ||
+                lhs.Depth != rhs.Depth ||
+                lhs.LinearLayout != rhs.LinearLayout ||
+                lhs.MemoryLayout.Packed != rhs.MemoryLayout.Packed)
+            {
+                return false;
+            }
+
+            if (lhs.LinearLayout)
+            {
+                return lhs.Stride == rhs.Stride;
+            }
+            else
+            {
+                return lhs.Width == rhs.Width;
+            }
+        }
+
+        /// <summary>
+        /// Determine if the given region covers the full texture, also considering width alignment.
+        /// </summary>
+        /// <param name="texture">The texture to check</param>
+        /// <param name="formatInfo"></param>
+        /// <param name="x1">Region start x</param>
+        /// <param name="y1">Region start y</param>
+        /// <param name="x2">Region end x</param>
+        /// <param name="y2">Region end y</param>
+        /// <returns>True if the region covers the full texture, false otherwise</returns>
+        private bool IsCopyRegionComplete(TwodTexture texture, FormatInfo formatInfo, int x1, int y1, int x2, int y2)
+        {
+            if (x1 != 0 || y1 != 0 || y2 != texture.Height)
+            {
+                return false;
+            }
+
+            int width;
+            int widthAlignment;
+
+            if (texture.LinearLayout)
+            {
+                widthAlignment = 1;
+                width = texture.Stride / formatInfo.BytesPerPixel;
+            }
+            else
+            {
+                widthAlignment = Constants.GobAlignment / formatInfo.BytesPerPixel;
+                width = texture.Width;
+            }
+
+            return width == BitUtils.AlignUp(x2, widthAlignment);
+        }
+
+        /// <summary>
+        /// Performs a full data copy between two textures, reading and writing guest memory directly.
+        /// The textures must have a matching layout, size, and bytes per pixel.
+        /// </summary>
+        /// <param name="src">The source texture</param>
+        /// <param name="dst">The destination texture</param>
+        /// <param name="w">Copy width</param>
+        /// <param name="h">Copy height</param>
+        /// <param name="bpp">Bytes per pixel</param>
+        private void UnscaledFullCopy(TwodTexture src, TwodTexture dst, int w, int h, int bpp)
+        {
+            var srcCalculator = new OffsetCalculator(
+                w,
+                h,
+                src.Stride,
+                src.LinearLayout,
+                src.MemoryLayout.UnpackGobBlocksInY(),
+                src.MemoryLayout.UnpackGobBlocksInZ(),
+                bpp);
+
+            (int _, int srcSize) = srcCalculator.GetRectangleRange(0, 0, w, h);
+
+            var memoryManager = _channel.MemoryManager;
+
+            ulong srcGpuVa = src.Address.Pack();
+            ulong dstGpuVa = dst.Address.Pack();
+
+            ReadOnlySpan<byte> srcSpan = memoryManager.GetSpan(srcGpuVa, srcSize, true);
+
+            int width;
+            int height = src.Height;
+            if (src.LinearLayout)
+            {
+                width = src.Stride / bpp;
+            }
+            else
+            {
+                width = src.Width;
+            }
+
+            // If the copy is not equal to the width and height of the texture, we will need to copy partially.
+            // It's worth noting that it has already been established that the src and dst are the same size.
+
+            if (w == width && h == height)
+            {
+                memoryManager.Write(dstGpuVa, srcSpan);
+            }
+            else
+            {
+                using WritableRegion dstRegion = memoryManager.GetWritableRegion(dstGpuVa, srcSize, true);
+                Span<byte> dstSpan = dstRegion.Memory.Span;
+
+                if (src.LinearLayout)
+                {
+                    int stride = src.Stride;
+                    int offset = 0;
+                    int lineSize = width * bpp;
+
+                    for (int y = 0; y < height; y++)
+                    {
+                        srcSpan.Slice(offset, lineSize).CopyTo(dstSpan.Slice(offset));
+
+                        offset += stride;
+                    }
+                }
+                else
+                {
+                    // Copy with the block linear layout in mind.
+                    // Recreate the offset calculate with bpp 1 for copy.
+
+                    int stride = w * bpp;
+
+                    srcCalculator = new OffsetCalculator(
+                        stride,
+                        h,
+                        0,
+                        false,
+                        src.MemoryLayout.UnpackGobBlocksInY(),
+                        src.MemoryLayout.UnpackGobBlocksInZ(),
+                        1);
+
+                    int strideTrunc = BitUtils.AlignDown(stride, 16);
+
+                    ReadOnlySpan<Vector128<byte>> srcVec = MemoryMarshal.Cast<byte, Vector128<byte>>(srcSpan);
+                    Span<Vector128<byte>> dstVec = MemoryMarshal.Cast<byte, Vector128<byte>>(dstSpan);
+
+                    for (int y = 0; y < h; y++)
+                    {
+                        int x = 0;
+
+                        srcCalculator.SetY(y);
+
+                        for (; x < strideTrunc; x += 16)
+                        {
+                            int offset = srcCalculator.GetOffset(x) >> 4;
+
+                            dstVec[offset] = srcVec[offset];
+                        }
+
+                        for (; x < stride; x++)
+                        {
+                            int offset = srcCalculator.GetOffset(x);
+
+                            dstSpan[offset] = srcSpan[offset];
+                        }
+                    }
+                }
+            }
+        }
+
+        /// <summary>
+        /// Performs the blit operation, triggered by the register write.
+        /// </summary>
+        /// <param name="argument">Method call argument</param>
+        private void PixelsFromMemorySrcY0Int(int argument)
+        {
+            var memoryManager = _channel.MemoryManager;
+
+            var dstCopyTexture = Unsafe.As<uint, TwodTexture>(ref _state.State.SetDstFormat);
+            var srcCopyTexture = Unsafe.As<uint, TwodTexture>(ref _state.State.SetSrcFormat);
+
+            long srcX = ((long)_state.State.SetPixelsFromMemorySrcX0Int << 32) | (long)(ulong)_state.State.SetPixelsFromMemorySrcX0Frac;
+            long srcY = ((long)_state.State.PixelsFromMemorySrcY0Int << 32) | (long)(ulong)_state.State.SetPixelsFromMemorySrcY0Frac;
+
+            long duDx = ((long)_state.State.SetPixelsFromMemoryDuDxInt << 32) | (long)(ulong)_state.State.SetPixelsFromMemoryDuDxFrac;
+            long dvDy = ((long)_state.State.SetPixelsFromMemoryDvDyInt << 32) | (long)(ulong)_state.State.SetPixelsFromMemoryDvDyFrac;
+
+            bool originCorner = _state.State.SetPixelsFromMemorySampleModeOrigin == SetPixelsFromMemorySampleModeOrigin.Corner;
+
+            if (originCorner)
+            {
+                // If the origin is corner, it is assumed that the guest API
+                // is manually centering the origin by adding a offset to the
+                // source region X/Y coordinates.
+                // Here we attempt to remove such offset to ensure we have the correct region.
+                // The offset is calculated as FactorXY / 2.0, where FactorXY = SrcXY / DstXY,
+                // so we do the same here by dividing the fixed point value by 2, while
+                // throwing away the fractional part to avoid rounding errors.
+                srcX -= (duDx >> 33) << 32;
+                srcY -= (dvDy >> 33) << 32;
+            }
+
+            int srcX1 = (int)(srcX >> 32);
+            int srcY1 = (int)(srcY >> 32);
+
+            int srcX2 = srcX1 + (int)((duDx * _state.State.SetPixelsFromMemoryDstWidth + uint.MaxValue) >> 32);
+            int srcY2 = srcY1 + (int)((dvDy * _state.State.SetPixelsFromMemoryDstHeight + uint.MaxValue) >> 32);
+
+            int dstX1 = (int)_state.State.SetPixelsFromMemoryDstX0;
+            int dstY1 = (int)_state.State.SetPixelsFromMemoryDstY0;
+
+            int dstX2 = dstX1 + (int)_state.State.SetPixelsFromMemoryDstWidth;
+            int dstY2 = dstY1 + (int)_state.State.SetPixelsFromMemoryDstHeight;
+
+            // The source and destination textures should at least be as big as the region being requested.
+            // The hints will only resize within alignment constraints, so out of bound copies won't resize in most cases.
+            var srcHint = new Size(srcX2, srcY2, 1);
+            var dstHint = new Size(dstX2, dstY2, 1);
+
+            var srcCopyTextureFormat = srcCopyTexture.Format.Convert();
+
+            int srcWidthAligned = srcCopyTexture.Stride / srcCopyTextureFormat.BytesPerPixel;
+
+            ulong offset = 0;
+
+            // For an out of bounds copy, we must ensure that the copy wraps to the next line,
+            // so for a copy from a 64x64 texture, in the region [32, 96[, there are 32 pixels that are
+            // outside the bounds of the texture. We fill the destination with the first 32 pixels
+            // of the next line on the source texture.
+            // This can be done by simply adding an offset to the texture address, so that the initial
+            // gap is skipped and the copy is inside bounds again.
+            // This is required by the proprietary guest OpenGL driver.
+            if (srcCopyTexture.LinearLayout && srcCopyTexture.Width == srcX2 && srcX2 > srcWidthAligned && srcX1 > 0)
+            {
+                offset = (ulong)(srcX1 * srcCopyTextureFormat.BytesPerPixel);
+                srcCopyTexture.Width -= srcX1;
+                srcX2 -= srcX1;
+                srcX1 = 0;
+            }
+
+            FormatInfo dstCopyTextureFormat = dstCopyTexture.Format.Convert();
+
+            bool canDirectCopy = GraphicsConfig.Fast2DCopy &&
+                srcX2 == dstX2 && srcY2 == dstY2 &&
+                IsDataCompatible(srcCopyTexture, dstCopyTexture, srcCopyTextureFormat, dstCopyTextureFormat) &&
+                IsCopyRegionComplete(srcCopyTexture, srcCopyTextureFormat, srcX1, srcY1, srcX2, srcY2) &&
+                IsCopyRegionComplete(dstCopyTexture, dstCopyTextureFormat, dstX1, dstY1, dstX2, dstY2);
+
+            var srcTexture = memoryManager.Physical.TextureCache.FindOrCreateTexture(
+                memoryManager,
+                srcCopyTexture,
+                offset,
+                srcCopyTextureFormat,
+                !canDirectCopy,
+                false,
+                srcHint);
+
+            if (srcTexture == null)
+            {
+                if (canDirectCopy)
+                {
+                    // Directly copy the data on CPU.
+                    UnscaledFullCopy(srcCopyTexture, dstCopyTexture, srcX2, srcY2, srcCopyTextureFormat.BytesPerPixel);
+                }
+
+                return;
+            }
+
+            memoryManager.Physical.TextureCache.Lift(srcTexture);
+
+            // When the source texture that was found has a depth format,
+            // we must enforce the target texture also has a depth format,
+            // as copies between depth and color formats are not allowed.
+
+            if (srcTexture.Format.IsDepthOrStencil())
+            {
+                dstCopyTextureFormat = srcTexture.Info.FormatInfo;
+            }
+            else
+            {
+                dstCopyTextureFormat = dstCopyTexture.Format.Convert();
+            }
+
+            var dstTexture = memoryManager.Physical.TextureCache.FindOrCreateTexture(
+                memoryManager,
+                dstCopyTexture,
+                0,
+                dstCopyTextureFormat,
+                true,
+                srcTexture.ScaleMode == TextureScaleMode.Scaled,
+                dstHint);
+
+            if (dstTexture == null)
+            {
+                return;
+            }
+
+            if (srcTexture.Info.Samples > 1 || dstTexture.Info.Samples > 1)
+            {
+                srcTexture.PropagateScale(dstTexture);
+            }
+
+            float scale = srcTexture.ScaleFactor;
+            float dstScale = dstTexture.ScaleFactor;
+
+            Extents2D srcRegion = new Extents2D(
+                (int)Math.Ceiling(scale * (srcX1 / srcTexture.Info.SamplesInX)),
+                (int)Math.Ceiling(scale * (srcY1 / srcTexture.Info.SamplesInY)),
+                (int)Math.Ceiling(scale * (srcX2 / srcTexture.Info.SamplesInX)),
+                (int)Math.Ceiling(scale * (srcY2 / srcTexture.Info.SamplesInY)));
+
+            Extents2D dstRegion = new Extents2D(
+                (int)Math.Ceiling(dstScale * (dstX1 / dstTexture.Info.SamplesInX)),
+                (int)Math.Ceiling(dstScale * (dstY1 / dstTexture.Info.SamplesInY)),
+                (int)Math.Ceiling(dstScale * (dstX2 / dstTexture.Info.SamplesInX)),
+                (int)Math.Ceiling(dstScale * (dstY2 / dstTexture.Info.SamplesInY)));
+
+            bool linearFilter = _state.State.SetPixelsFromMemorySampleModeFilter == SetPixelsFromMemorySampleModeFilter.Bilinear;
+
+            srcTexture.HostTexture.CopyTo(dstTexture.HostTexture, srcRegion, dstRegion, linearFilter);
+
+            dstTexture.SignalModified();
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Twod/TwodClassState.cs b/src/Ryujinx.Graphics.Gpu/Engine/Twod/TwodClassState.cs
new file mode 100644
index 00000000..46fddb04
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Twod/TwodClassState.cs
@@ -0,0 +1,816 @@
+// This file was auto-generated from NVIDIA official Maxwell definitions.
+
+using Ryujinx.Common.Memory;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Twod
+{
+    /// <summary>
+    /// Notify type.
+    /// </summary>
+    enum NotifyType
+    {
+        WriteOnly = 0,
+        WriteThenAwaken = 1,
+    }
+
+    /// <summary>
+    /// Format of the destination texture.
+    /// </summary>
+    enum SetDstFormatV
+    {
+        A8r8g8b8 = 207,
+        A8rl8gl8bl8 = 208,
+        A2r10g10b10 = 223,
+        A8b8g8r8 = 213,
+        A8bl8gl8rl8 = 214,
+        A2b10g10r10 = 209,
+        X8r8g8b8 = 230,
+        X8rl8gl8bl8 = 231,
+        X8b8g8r8 = 249,
+        X8bl8gl8rl8 = 250,
+        R5g6b5 = 232,
+        A1r5g5b5 = 233,
+        X1r5g5b5 = 248,
+        Y8 = 243,
+        Y16 = 238,
+        Y32 = 255,
+        Z1r5g5b5 = 251,
+        O1r5g5b5 = 252,
+        Z8r8g8b8 = 253,
+        O8r8g8b8 = 254,
+        Y18x8 = 28,
+        Rf16 = 242,
+        Rf32 = 229,
+        Rf32Gf32 = 203,
+        Rf16Gf16Bf16Af16 = 202,
+        Rf16Gf16Bf16X16 = 206,
+        Rf32Gf32Bf32Af32 = 192,
+        Rf32Gf32Bf32X32 = 195,
+        R16G16B16A16 = 198,
+        Rn16Gn16Bn16An16 = 199,
+        Bf10gf11rf11 = 224,
+        An8bn8gn8rn8 = 215,
+        Rf16Gf16 = 222,
+        R16G16 = 218,
+        Rn16Gn16 = 219,
+        G8r8 = 234,
+        Gn8rn8 = 235,
+        Rn16 = 239,
+        Rn8 = 244,
+        A8 = 247,
+    }
+
+    /// <summary>
+    /// Memory layout of the destination texture.
+    /// </summary>
+    enum SetDstMemoryLayoutV
+    {
+        Blocklinear = 0,
+        Pitch = 1,
+    }
+
+    /// <summary>
+    /// Height in GOBs of the destination texture.
+    /// </summary>
+    enum SetDstBlockSizeHeight
+    {
+        OneGob = 0,
+        TwoGobs = 1,
+        FourGobs = 2,
+        EightGobs = 3,
+        SixteenGobs = 4,
+        ThirtytwoGobs = 5,
+    }
+
+    /// <summary>
+    /// Depth in GOBs of the destination texture.
+    /// </summary>
+    enum SetDstBlockSizeDepth
+    {
+        OneGob = 0,
+        TwoGobs = 1,
+        FourGobs = 2,
+        EightGobs = 3,
+        SixteenGobs = 4,
+        ThirtytwoGobs = 5,
+    }
+
+    /// <summary>
+    /// Format of the source texture.
+    /// </summary>
+    enum SetSrcFormatV
+    {
+        A8r8g8b8 = 207,
+        A8rl8gl8bl8 = 208,
+        A2r10g10b10 = 223,
+        A8b8g8r8 = 213,
+        A8bl8gl8rl8 = 214,
+        A2b10g10r10 = 209,
+        X8r8g8b8 = 230,
+        X8rl8gl8bl8 = 231,
+        X8b8g8r8 = 249,
+        X8bl8gl8rl8 = 250,
+        R5g6b5 = 232,
+        A1r5g5b5 = 233,
+        X1r5g5b5 = 248,
+        Y8 = 243,
+        Ay8 = 29,
+        Y16 = 238,
+        Y32 = 255,
+        Z1r5g5b5 = 251,
+        O1r5g5b5 = 252,
+        Z8r8g8b8 = 253,
+        O8r8g8b8 = 254,
+        Y18x8 = 28,
+        Rf16 = 242,
+        Rf32 = 229,
+        Rf32Gf32 = 203,
+        Rf16Gf16Bf16Af16 = 202,
+        Rf16Gf16Bf16X16 = 206,
+        Rf32Gf32Bf32Af32 = 192,
+        Rf32Gf32Bf32X32 = 195,
+        R16G16B16A16 = 198,
+        Rn16Gn16Bn16An16 = 199,
+        Bf10gf11rf11 = 224,
+        An8bn8gn8rn8 = 215,
+        Rf16Gf16 = 222,
+        R16G16 = 218,
+        Rn16Gn16 = 219,
+        G8r8 = 234,
+        Gn8rn8 = 235,
+        Rn16 = 239,
+        Rn8 = 244,
+        A8 = 247,
+    }
+
+    /// <summary>
+    /// Memory layout of the source texture.
+    /// </summary>
+    enum SetSrcMemoryLayoutV
+    {
+        Blocklinear = 0,
+        Pitch = 1,
+    }
+
+    /// <summary>
+    /// Height in GOBs of the source texture.
+    /// </summary>
+    enum SetSrcBlockSizeHeight
+    {
+        OneGob = 0,
+        TwoGobs = 1,
+        FourGobs = 2,
+        EightGobs = 3,
+        SixteenGobs = 4,
+        ThirtytwoGobs = 5,
+    }
+
+    /// <summary>
+    /// Depth in GOBs of the source texture.
+    /// </summary>
+    enum SetSrcBlockSizeDepth
+    {
+        OneGob = 0,
+        TwoGobs = 1,
+        FourGobs = 2,
+        EightGobs = 3,
+        SixteenGobs = 4,
+        ThirtytwoGobs = 5,
+    }
+
+    /// <summary>
+    /// Texture data caches to invalidate.
+    /// </summary>
+    enum TwodInvalidateTextureDataCacheV
+    {
+        L1Only = 0,
+        L2Only = 1,
+        L1AndL2 = 2,
+    }
+
+    /// <summary>
+    /// Sector promotion parameters.
+    /// </summary>
+    enum SetPixelsFromMemorySectorPromotionV
+    {
+        NoPromotion = 0,
+        PromoteTo2V = 1,
+        PromoteTo2H = 2,
+        PromoteTo4 = 3,
+    }
+
+    /// <summary>
+    /// Number of processing clusters.
+    /// </summary>
+    enum SetNumProcessingClustersV
+    {
+        All = 0,
+        One = 1,
+    }
+
+    /// <summary>
+    /// Color key format.
+    /// </summary>
+    enum SetColorKeyFormatV
+    {
+        A16r5g6b5 = 0,
+        A1r5g5b5 = 1,
+        A8r8g8b8 = 2,
+        A2r10g10b10 = 3,
+        Y8 = 4,
+        Y16 = 5,
+        Y32 = 6,
+    }
+
+    /// <summary>
+    /// Color blit operation.
+    /// </summary>
+    enum SetOperationV
+    {
+        SrccopyAnd = 0,
+        RopAnd = 1,
+        BlendAnd = 2,
+        Srccopy = 3,
+        Rop = 4,
+        SrccopyPremult = 5,
+        BlendPremult = 6,
+    }
+
+    /// <summary>
+    /// Texture pattern selection.
+    /// </summary>
+    enum SetPatternSelectV
+    {
+        Monochrome8x8 = 0,
+        Monochrome64x1 = 1,
+        Monochrome1x64 = 2,
+        Color = 3,
+    }
+
+    /// <summary>
+    /// Render enable override mode.
+    /// </summary>
+    enum SetRenderEnableOverrideMode
+    {
+        UseRenderEnable = 0,
+        AlwaysRender = 1,
+        NeverRender = 2,
+    }
+
+    /// <summary>
+    /// Pixels from memory horizontal direction.
+    /// </summary>
+    enum SetPixelsFromMemoryDirectionHorizontal
+    {
+        HwDecides = 0,
+        LeftToRight = 1,
+        RightToLeft = 2,
+    }
+
+    /// <summary>
+    /// Pixels from memory vertical direction.
+    /// </summary>
+    enum SetPixelsFromMemoryDirectionVertical
+    {
+        HwDecides = 0,
+        TopToBottom = 1,
+        BottomToTop = 2,
+    }
+
+    /// <summary>
+    /// Color format of the monochrome pattern.
+    /// </summary>
+    enum SetMonochromePatternColorFormatV
+    {
+        A8x8r5g6b5 = 0,
+        A1r5g5b5 = 1,
+        A8r8g8b8 = 2,
+        A8y8 = 3,
+        A8x8y16 = 4,
+        Y32 = 5,
+        ByteExpand = 6,
+    }
+
+    /// <summary>
+    /// Format of the monochrome pattern.
+    /// </summary>
+    enum SetMonochromePatternFormatV
+    {
+        Cga6M1 = 0,
+        LeM1 = 1,
+    }
+
+    /// <summary>
+    /// DMA semaphore reduction operation.
+    /// </summary>
+    enum MmeDmaReductionReductionOp
+    {
+        RedAdd = 0,
+        RedMin = 1,
+        RedMax = 2,
+        RedInc = 3,
+        RedDec = 4,
+        RedAnd = 5,
+        RedOr = 6,
+        RedXor = 7,
+    }
+
+    /// <summary>
+    /// DMA semaphore reduction format.
+    /// </summary>
+    enum MmeDmaReductionReductionFormat
+    {
+        Unsigned = 0,
+        Signed = 1,
+    }
+
+    /// <summary>
+    /// DMA semaphore reduction size.
+    /// </summary>
+    enum MmeDmaReductionReductionSize
+    {
+        FourBytes = 0,
+        EightBytes = 1,
+    }
+
+    /// <summary>
+    /// Data FIFO size.
+    /// </summary>
+    enum SetMmeDataFifoConfigFifoSize
+    {
+        Size0kb = 0,
+        Size4kb = 1,
+        Size8kb = 2,
+        Size12kb = 3,
+        Size16kb = 4,
+    }
+
+    /// <summary>
+    /// Render solid primitive mode.
+    /// </summary>
+    enum RenderSolidPrimModeV
+    {
+        Points = 0,
+        Lines = 1,
+        Polyline = 2,
+        Triangles = 3,
+        Rects = 4,
+    }
+
+    /// <summary>
+    /// Render solid primitive color format.
+    /// </summary>
+    enum SetRenderSolidPrimColorFormatV
+    {
+        Rf32Gf32Bf32Af32 = 192,
+        Rf16Gf16Bf16Af16 = 202,
+        Rf32Gf32 = 203,
+        A8r8g8b8 = 207,
+        A2r10g10b10 = 223,
+        A8b8g8r8 = 213,
+        A2b10g10r10 = 209,
+        X8r8g8b8 = 230,
+        X8b8g8r8 = 249,
+        R5g6b5 = 232,
+        A1r5g5b5 = 233,
+        X1r5g5b5 = 248,
+        Y8 = 243,
+        Y16 = 238,
+        Y32 = 255,
+        Z1r5g5b5 = 251,
+        O1r5g5b5 = 252,
+        Z8r8g8b8 = 253,
+        O8r8g8b8 = 254,
+    }
+
+    /// <summary>
+    /// Pixels from CPU data type.
+    /// </summary>
+    enum SetPixelsFromCpuDataTypeV
+    {
+        Color = 0,
+        Index = 1,
+    }
+
+    /// <summary>
+    /// Pixels from CPU color format.
+    /// </summary>
+    enum SetPixelsFromCpuColorFormatV
+    {
+        A8r8g8b8 = 207,
+        A2r10g10b10 = 223,
+        A8b8g8r8 = 213,
+        A2b10g10r10 = 209,
+        X8r8g8b8 = 230,
+        X8b8g8r8 = 249,
+        R5g6b5 = 232,
+        A1r5g5b5 = 233,
+        X1r5g5b5 = 248,
+        Y8 = 243,
+        Y16 = 238,
+        Y32 = 255,
+        Z1r5g5b5 = 251,
+        O1r5g5b5 = 252,
+        Z8r8g8b8 = 253,
+        O8r8g8b8 = 254,
+    }
+
+    /// <summary>
+    /// Pixels from CPU palette index format.
+    /// </summary>
+    enum SetPixelsFromCpuIndexFormatV
+    {
+        I1 = 0,
+        I4 = 1,
+        I8 = 2,
+    }
+
+    /// <summary>
+    /// Pixels from CPU monochrome format.
+    /// </summary>
+    enum SetPixelsFromCpuMonoFormatV
+    {
+        Cga6M1 = 0,
+        LeM1 = 1,
+    }
+
+    /// <summary>
+    /// Pixels from CPU wrap mode.
+    /// </summary>
+    enum SetPixelsFromCpuWrapV
+    {
+        WrapPixel = 0,
+        WrapByte = 1,
+        WrapDword = 2,
+    }
+
+    /// <summary>
+    /// Pixels from CPU monochrome opacity.
+    /// </summary>
+    enum SetPixelsFromCpuMonoOpacityV
+    {
+        Transparent = 0,
+        Opaque = 1,
+    }
+
+    /// <summary>
+    /// Pixels from memory block shape.
+    /// </summary>
+    enum SetPixelsFromMemoryBlockShapeV
+    {
+        Auto = 0,
+        Shape8x8 = 1,
+        Shape16x4 = 2,
+    }
+
+    /// <summary>
+    /// Pixels from memory origin.
+    /// </summary>
+    enum SetPixelsFromMemorySampleModeOrigin
+    {
+        Center = 0,
+        Corner = 1,
+    }
+
+    /// <summary>
+    /// Pixels from memory filter mode.
+    /// </summary>
+    enum SetPixelsFromMemorySampleModeFilter
+    {
+        Point = 0,
+        Bilinear = 1,
+    }
+
+    /// <summary>
+    /// Render solid primitive point coordinates.
+    /// </summary>
+    struct RenderSolidPrimPoint
+    {
+#pragma warning disable CS0649
+        public uint SetX;
+        public uint Y;
+#pragma warning restore CS0649
+    }
+
+    /// <summary>
+    /// 2D class state.
+    /// </summary>
+    unsafe struct TwodClassState : 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) & 0x1FFFFFF);
+        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 fixed uint Reserved144[3];
+        public uint SetInstrumentationMethodHeader;
+        public uint SetInstrumentationMethodData;
+        public fixed uint Reserved158[37];
+        public uint SetMmeSwitchState;
+        public bool SetMmeSwitchStateValid => (SetMmeSwitchState & 0x1) != 0;
+        public int SetMmeSwitchStateSaveMacro => (int)((SetMmeSwitchState >> 4) & 0xFF);
+        public int SetMmeSwitchStateRestoreMacro => (int)((SetMmeSwitchState >> 12) & 0xFF);
+        public fixed uint Reserved1F0[4];
+        public uint SetDstFormat;
+        public SetDstFormatV SetDstFormatV => (SetDstFormatV)((SetDstFormat >> 0) & 0xFF);
+        public uint SetDstMemoryLayout;
+        public SetDstMemoryLayoutV SetDstMemoryLayoutV => (SetDstMemoryLayoutV)((SetDstMemoryLayout >> 0) & 0x1);
+        public uint SetDstBlockSize;
+        public SetDstBlockSizeHeight SetDstBlockSizeHeight => (SetDstBlockSizeHeight)((SetDstBlockSize >> 4) & 0x7);
+        public SetDstBlockSizeDepth SetDstBlockSizeDepth => (SetDstBlockSizeDepth)((SetDstBlockSize >> 8) & 0x7);
+        public uint SetDstDepth;
+        public uint SetDstLayer;
+        public uint SetDstPitch;
+        public uint SetDstWidth;
+        public uint SetDstHeight;
+        public uint SetDstOffsetUpper;
+        public int SetDstOffsetUpperV => (int)((SetDstOffsetUpper >> 0) & 0xFF);
+        public uint SetDstOffsetLower;
+        public uint FlushAndInvalidateRopMiniCache;
+        public bool FlushAndInvalidateRopMiniCacheV => (FlushAndInvalidateRopMiniCache & 0x1) != 0;
+        public uint SetSpareNoop06;
+        public uint SetSrcFormat;
+        public SetSrcFormatV SetSrcFormatV => (SetSrcFormatV)((SetSrcFormat >> 0) & 0xFF);
+        public uint SetSrcMemoryLayout;
+        public SetSrcMemoryLayoutV SetSrcMemoryLayoutV => (SetSrcMemoryLayoutV)((SetSrcMemoryLayout >> 0) & 0x1);
+        public uint SetSrcBlockSize;
+        public SetSrcBlockSizeHeight SetSrcBlockSizeHeight => (SetSrcBlockSizeHeight)((SetSrcBlockSize >> 4) & 0x7);
+        public SetSrcBlockSizeDepth SetSrcBlockSizeDepth => (SetSrcBlockSizeDepth)((SetSrcBlockSize >> 8) & 0x7);
+        public uint SetSrcDepth;
+        public uint TwodInvalidateTextureDataCache;
+        public TwodInvalidateTextureDataCacheV TwodInvalidateTextureDataCacheV => (TwodInvalidateTextureDataCacheV)((TwodInvalidateTextureDataCache >> 0) & 0x3);
+        public uint SetSrcPitch;
+        public uint SetSrcWidth;
+        public uint SetSrcHeight;
+        public uint SetSrcOffsetUpper;
+        public int SetSrcOffsetUpperV => (int)((SetSrcOffsetUpper >> 0) & 0xFF);
+        public uint SetSrcOffsetLower;
+        public uint SetPixelsFromMemorySectorPromotion;
+        public SetPixelsFromMemorySectorPromotionV SetPixelsFromMemorySectorPromotionV => (SetPixelsFromMemorySectorPromotionV)((SetPixelsFromMemorySectorPromotion >> 0) & 0x3);
+        public uint SetSpareNoop12;
+        public uint SetNumProcessingClusters;
+        public SetNumProcessingClustersV SetNumProcessingClustersV => (SetNumProcessingClustersV)((SetNumProcessingClusters >> 0) & 0x1);
+        public uint SetRenderEnableA;
+        public int SetRenderEnableAOffsetUpper => (int)((SetRenderEnableA >> 0) & 0xFF);
+        public uint SetRenderEnableB;
+        public uint SetRenderEnableC;
+        public int SetRenderEnableCMode => (int)((SetRenderEnableC >> 0) & 0x7);
+        public uint SetSpareNoop08;
+        public uint SetSpareNoop01;
+        public uint SetSpareNoop11;
+        public uint SetSpareNoop07;
+        public uint SetClipX0;
+        public uint SetClipY0;
+        public uint SetClipWidth;
+        public uint SetClipHeight;
+        public uint SetClipEnable;
+        public bool SetClipEnableV => (SetClipEnable & 0x1) != 0;
+        public uint SetColorKeyFormat;
+        public SetColorKeyFormatV SetColorKeyFormatV => (SetColorKeyFormatV)((SetColorKeyFormat >> 0) & 0x7);
+        public uint SetColorKey;
+        public uint SetColorKeyEnable;
+        public bool SetColorKeyEnableV => (SetColorKeyEnable & 0x1) != 0;
+        public uint SetRop;
+        public int SetRopV => (int)((SetRop >> 0) & 0xFF);
+        public uint SetBeta1;
+        public uint SetBeta4;
+        public int SetBeta4B => (int)((SetBeta4 >> 0) & 0xFF);
+        public int SetBeta4G => (int)((SetBeta4 >> 8) & 0xFF);
+        public int SetBeta4R => (int)((SetBeta4 >> 16) & 0xFF);
+        public int SetBeta4A => (int)((SetBeta4 >> 24) & 0xFF);
+        public uint SetOperation;
+        public SetOperationV SetOperationV => (SetOperationV)((SetOperation >> 0) & 0x7);
+        public uint SetPatternOffset;
+        public int SetPatternOffsetX => (int)((SetPatternOffset >> 0) & 0x3F);
+        public int SetPatternOffsetY => (int)((SetPatternOffset >> 8) & 0x3F);
+        public uint SetPatternSelect;
+        public SetPatternSelectV SetPatternSelectV => (SetPatternSelectV)((SetPatternSelect >> 0) & 0x3);
+        public uint SetDstColorRenderToZetaSurface;
+        public bool SetDstColorRenderToZetaSurfaceV => (SetDstColorRenderToZetaSurface & 0x1) != 0;
+        public uint SetSpareNoop04;
+        public uint SetSpareNoop15;
+        public uint SetSpareNoop13;
+        public uint SetSpareNoop03;
+        public uint SetSpareNoop14;
+        public uint SetSpareNoop02;
+        public uint SetCompression;
+        public bool SetCompressionEnable => (SetCompression & 0x1) != 0;
+        public uint SetSpareNoop09;
+        public uint SetRenderEnableOverride;
+        public SetRenderEnableOverrideMode SetRenderEnableOverrideMode => (SetRenderEnableOverrideMode)((SetRenderEnableOverride >> 0) & 0x3);
+        public uint SetPixelsFromMemoryDirection;
+        public SetPixelsFromMemoryDirectionHorizontal SetPixelsFromMemoryDirectionHorizontal => (SetPixelsFromMemoryDirectionHorizontal)((SetPixelsFromMemoryDirection >> 0) & 0x3);
+        public SetPixelsFromMemoryDirectionVertical SetPixelsFromMemoryDirectionVertical => (SetPixelsFromMemoryDirectionVertical)((SetPixelsFromMemoryDirection >> 4) & 0x3);
+        public uint SetSpareNoop10;
+        public uint SetMonochromePatternColorFormat;
+        public SetMonochromePatternColorFormatV SetMonochromePatternColorFormatV => (SetMonochromePatternColorFormatV)((SetMonochromePatternColorFormat >> 0) & 0x7);
+        public uint SetMonochromePatternFormat;
+        public SetMonochromePatternFormatV SetMonochromePatternFormatV => (SetMonochromePatternFormatV)((SetMonochromePatternFormat >> 0) & 0x1);
+        public uint SetMonochromePatternColor0;
+        public uint SetMonochromePatternColor1;
+        public uint SetMonochromePattern0;
+        public uint SetMonochromePattern1;
+        public Array64<uint> ColorPatternX8r8g8b8;
+        public int ColorPatternX8r8g8b8B0(int i) => (int)((ColorPatternX8r8g8b8[i] >> 0) & 0xFF);
+        public int ColorPatternX8r8g8b8G0(int i) => (int)((ColorPatternX8r8g8b8[i] >> 8) & 0xFF);
+        public int ColorPatternX8r8g8b8R0(int i) => (int)((ColorPatternX8r8g8b8[i] >> 16) & 0xFF);
+        public int ColorPatternX8r8g8b8Ignore0(int i) => (int)((ColorPatternX8r8g8b8[i] >> 24) & 0xFF);
+        public Array32<uint> ColorPatternR5g6b5;
+        public int ColorPatternR5g6b5B0(int i) => (int)((ColorPatternR5g6b5[i] >> 0) & 0x1F);
+        public int ColorPatternR5g6b5G0(int i) => (int)((ColorPatternR5g6b5[i] >> 5) & 0x3F);
+        public int ColorPatternR5g6b5R0(int i) => (int)((ColorPatternR5g6b5[i] >> 11) & 0x1F);
+        public int ColorPatternR5g6b5B1(int i) => (int)((ColorPatternR5g6b5[i] >> 16) & 0x1F);
+        public int ColorPatternR5g6b5G1(int i) => (int)((ColorPatternR5g6b5[i] >> 21) & 0x3F);
+        public int ColorPatternR5g6b5R1(int i) => (int)((ColorPatternR5g6b5[i] >> 27) & 0x1F);
+        public Array32<uint> ColorPatternX1r5g5b5;
+        public int ColorPatternX1r5g5b5B0(int i) => (int)((ColorPatternX1r5g5b5[i] >> 0) & 0x1F);
+        public int ColorPatternX1r5g5b5G0(int i) => (int)((ColorPatternX1r5g5b5[i] >> 5) & 0x1F);
+        public int ColorPatternX1r5g5b5R0(int i) => (int)((ColorPatternX1r5g5b5[i] >> 10) & 0x1F);
+        public bool ColorPatternX1r5g5b5Ignore0(int i) => (ColorPatternX1r5g5b5[i] & 0x8000) != 0;
+        public int ColorPatternX1r5g5b5B1(int i) => (int)((ColorPatternX1r5g5b5[i] >> 16) & 0x1F);
+        public int ColorPatternX1r5g5b5G1(int i) => (int)((ColorPatternX1r5g5b5[i] >> 21) & 0x1F);
+        public int ColorPatternX1r5g5b5R1(int i) => (int)((ColorPatternX1r5g5b5[i] >> 26) & 0x1F);
+        public bool ColorPatternX1r5g5b5Ignore1(int i) => (ColorPatternX1r5g5b5[i] & 0x80000000) != 0;
+        public Array16<uint> ColorPatternY8;
+        public int ColorPatternY8Y0(int i) => (int)((ColorPatternY8[i] >> 0) & 0xFF);
+        public int ColorPatternY8Y1(int i) => (int)((ColorPatternY8[i] >> 8) & 0xFF);
+        public int ColorPatternY8Y2(int i) => (int)((ColorPatternY8[i] >> 16) & 0xFF);
+        public int ColorPatternY8Y3(int i) => (int)((ColorPatternY8[i] >> 24) & 0xFF);
+        public uint SetRenderSolidPrimColor0;
+        public uint SetRenderSolidPrimColor1;
+        public uint SetRenderSolidPrimColor2;
+        public uint SetRenderSolidPrimColor3;
+        public uint SetMmeMemAddressA;
+        public int SetMmeMemAddressAUpper => (int)((SetMmeMemAddressA >> 0) & 0x1FFFFFF);
+        public uint SetMmeMemAddressB;
+        public uint SetMmeDataRamAddress;
+        public uint MmeDmaRead;
+        public uint MmeDmaReadFifoed;
+        public uint MmeDmaWrite;
+        public uint MmeDmaReduction;
+        public MmeDmaReductionReductionOp MmeDmaReductionReductionOp => (MmeDmaReductionReductionOp)((MmeDmaReduction >> 0) & 0x7);
+        public MmeDmaReductionReductionFormat MmeDmaReductionReductionFormat => (MmeDmaReductionReductionFormat)((MmeDmaReduction >> 4) & 0x3);
+        public MmeDmaReductionReductionSize MmeDmaReductionReductionSize => (MmeDmaReductionReductionSize)((MmeDmaReduction >> 8) & 0x1);
+        public uint MmeDmaSysmembar;
+        public bool MmeDmaSysmembarV => (MmeDmaSysmembar & 0x1) != 0;
+        public uint MmeDmaSync;
+        public uint SetMmeDataFifoConfig;
+        public SetMmeDataFifoConfigFifoSize SetMmeDataFifoConfigFifoSize => (SetMmeDataFifoConfigFifoSize)((SetMmeDataFifoConfig >> 0) & 0x7);
+        public fixed uint Reserved578[2];
+        public uint RenderSolidPrimMode;
+        public RenderSolidPrimModeV RenderSolidPrimModeV => (RenderSolidPrimModeV)((RenderSolidPrimMode >> 0) & 0x7);
+        public uint SetRenderSolidPrimColorFormat;
+        public SetRenderSolidPrimColorFormatV SetRenderSolidPrimColorFormatV => (SetRenderSolidPrimColorFormatV)((SetRenderSolidPrimColorFormat >> 0) & 0xFF);
+        public uint SetRenderSolidPrimColor;
+        public uint SetRenderSolidLineTieBreakBits;
+        public bool SetRenderSolidLineTieBreakBitsXmajXincYinc => (SetRenderSolidLineTieBreakBits & 0x1) != 0;
+        public bool SetRenderSolidLineTieBreakBitsXmajXdecYinc => (SetRenderSolidLineTieBreakBits & 0x10) != 0;
+        public bool SetRenderSolidLineTieBreakBitsYmajXincYinc => (SetRenderSolidLineTieBreakBits & 0x100) != 0;
+        public bool SetRenderSolidLineTieBreakBitsYmajXdecYinc => (SetRenderSolidLineTieBreakBits & 0x1000) != 0;
+        public fixed uint Reserved590[20];
+        public uint RenderSolidPrimPointXY;
+        public int RenderSolidPrimPointXYX => (int)((RenderSolidPrimPointXY >> 0) & 0xFFFF);
+        public int RenderSolidPrimPointXYY => (int)((RenderSolidPrimPointXY >> 16) & 0xFFFF);
+        public fixed uint Reserved5E4[7];
+        public Array64<RenderSolidPrimPoint> RenderSolidPrimPoint;
+        public uint SetPixelsFromCpuDataType;
+        public SetPixelsFromCpuDataTypeV SetPixelsFromCpuDataTypeV => (SetPixelsFromCpuDataTypeV)((SetPixelsFromCpuDataType >> 0) & 0x1);
+        public uint SetPixelsFromCpuColorFormat;
+        public SetPixelsFromCpuColorFormatV SetPixelsFromCpuColorFormatV => (SetPixelsFromCpuColorFormatV)((SetPixelsFromCpuColorFormat >> 0) & 0xFF);
+        public uint SetPixelsFromCpuIndexFormat;
+        public SetPixelsFromCpuIndexFormatV SetPixelsFromCpuIndexFormatV => (SetPixelsFromCpuIndexFormatV)((SetPixelsFromCpuIndexFormat >> 0) & 0x3);
+        public uint SetPixelsFromCpuMonoFormat;
+        public SetPixelsFromCpuMonoFormatV SetPixelsFromCpuMonoFormatV => (SetPixelsFromCpuMonoFormatV)((SetPixelsFromCpuMonoFormat >> 0) & 0x1);
+        public uint SetPixelsFromCpuWrap;
+        public SetPixelsFromCpuWrapV SetPixelsFromCpuWrapV => (SetPixelsFromCpuWrapV)((SetPixelsFromCpuWrap >> 0) & 0x3);
+        public uint SetPixelsFromCpuColor0;
+        public uint SetPixelsFromCpuColor1;
+        public uint SetPixelsFromCpuMonoOpacity;
+        public SetPixelsFromCpuMonoOpacityV SetPixelsFromCpuMonoOpacityV => (SetPixelsFromCpuMonoOpacityV)((SetPixelsFromCpuMonoOpacity >> 0) & 0x1);
+        public fixed uint Reserved820[6];
+        public uint SetPixelsFromCpuSrcWidth;
+        public uint SetPixelsFromCpuSrcHeight;
+        public uint SetPixelsFromCpuDxDuFrac;
+        public uint SetPixelsFromCpuDxDuInt;
+        public uint SetPixelsFromCpuDyDvFrac;
+        public uint SetPixelsFromCpuDyDvInt;
+        public uint SetPixelsFromCpuDstX0Frac;
+        public uint SetPixelsFromCpuDstX0Int;
+        public uint SetPixelsFromCpuDstY0Frac;
+        public uint SetPixelsFromCpuDstY0Int;
+        public uint PixelsFromCpuData;
+        public fixed uint Reserved864[3];
+        public uint SetBigEndianControl;
+        public bool SetBigEndianControlX32Swap1 => (SetBigEndianControl & 0x1) != 0;
+        public bool SetBigEndianControlX32Swap4 => (SetBigEndianControl & 0x2) != 0;
+        public bool SetBigEndianControlX32Swap8 => (SetBigEndianControl & 0x4) != 0;
+        public bool SetBigEndianControlX32Swap16 => (SetBigEndianControl & 0x8) != 0;
+        public bool SetBigEndianControlX16Swap1 => (SetBigEndianControl & 0x10) != 0;
+        public bool SetBigEndianControlX16Swap4 => (SetBigEndianControl & 0x20) != 0;
+        public bool SetBigEndianControlX16Swap8 => (SetBigEndianControl & 0x40) != 0;
+        public bool SetBigEndianControlX16Swap16 => (SetBigEndianControl & 0x80) != 0;
+        public bool SetBigEndianControlX8Swap1 => (SetBigEndianControl & 0x100) != 0;
+        public bool SetBigEndianControlX8Swap4 => (SetBigEndianControl & 0x200) != 0;
+        public bool SetBigEndianControlX8Swap8 => (SetBigEndianControl & 0x400) != 0;
+        public bool SetBigEndianControlX8Swap16 => (SetBigEndianControl & 0x800) != 0;
+        public bool SetBigEndianControlI1X8Cga6Swap1 => (SetBigEndianControl & 0x1000) != 0;
+        public bool SetBigEndianControlI1X8Cga6Swap4 => (SetBigEndianControl & 0x2000) != 0;
+        public bool SetBigEndianControlI1X8Cga6Swap8 => (SetBigEndianControl & 0x4000) != 0;
+        public bool SetBigEndianControlI1X8Cga6Swap16 => (SetBigEndianControl & 0x8000) != 0;
+        public bool SetBigEndianControlI1X8LeSwap1 => (SetBigEndianControl & 0x10000) != 0;
+        public bool SetBigEndianControlI1X8LeSwap4 => (SetBigEndianControl & 0x20000) != 0;
+        public bool SetBigEndianControlI1X8LeSwap8 => (SetBigEndianControl & 0x40000) != 0;
+        public bool SetBigEndianControlI1X8LeSwap16 => (SetBigEndianControl & 0x80000) != 0;
+        public bool SetBigEndianControlI4Swap1 => (SetBigEndianControl & 0x100000) != 0;
+        public bool SetBigEndianControlI4Swap4 => (SetBigEndianControl & 0x200000) != 0;
+        public bool SetBigEndianControlI4Swap8 => (SetBigEndianControl & 0x400000) != 0;
+        public bool SetBigEndianControlI4Swap16 => (SetBigEndianControl & 0x800000) != 0;
+        public bool SetBigEndianControlI8Swap1 => (SetBigEndianControl & 0x1000000) != 0;
+        public bool SetBigEndianControlI8Swap4 => (SetBigEndianControl & 0x2000000) != 0;
+        public bool SetBigEndianControlI8Swap8 => (SetBigEndianControl & 0x4000000) != 0;
+        public bool SetBigEndianControlI8Swap16 => (SetBigEndianControl & 0x8000000) != 0;
+        public bool SetBigEndianControlOverride => (SetBigEndianControl & 0x10000000) != 0;
+        public fixed uint Reserved874[3];
+        public uint SetPixelsFromMemoryBlockShape;
+        public SetPixelsFromMemoryBlockShapeV SetPixelsFromMemoryBlockShapeV => (SetPixelsFromMemoryBlockShapeV)((SetPixelsFromMemoryBlockShape >> 0) & 0x7);
+        public uint SetPixelsFromMemoryCorralSize;
+        public int SetPixelsFromMemoryCorralSizeV => (int)((SetPixelsFromMemoryCorralSize >> 0) & 0x3FF);
+        public uint SetPixelsFromMemorySafeOverlap;
+        public bool SetPixelsFromMemorySafeOverlapV => (SetPixelsFromMemorySafeOverlap & 0x1) != 0;
+        public uint SetPixelsFromMemorySampleMode;
+        public SetPixelsFromMemorySampleModeOrigin SetPixelsFromMemorySampleModeOrigin => (SetPixelsFromMemorySampleModeOrigin)((SetPixelsFromMemorySampleMode >> 0) & 0x1);
+        public SetPixelsFromMemorySampleModeFilter SetPixelsFromMemorySampleModeFilter => (SetPixelsFromMemorySampleModeFilter)((SetPixelsFromMemorySampleMode >> 4) & 0x1);
+        public fixed uint Reserved890[8];
+        public uint SetPixelsFromMemoryDstX0;
+        public uint SetPixelsFromMemoryDstY0;
+        public uint SetPixelsFromMemoryDstWidth;
+        public uint SetPixelsFromMemoryDstHeight;
+        public uint SetPixelsFromMemoryDuDxFrac;
+        public uint SetPixelsFromMemoryDuDxInt;
+        public uint SetPixelsFromMemoryDvDyFrac;
+        public uint SetPixelsFromMemoryDvDyInt;
+        public uint SetPixelsFromMemorySrcX0Frac;
+        public uint SetPixelsFromMemorySrcX0Int;
+        public uint SetPixelsFromMemorySrcY0Frac;
+        public uint PixelsFromMemorySrcY0Int;
+        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 fixed uint Reserved960[291];
+        public uint MmeDmaWriteMethodBarrier;
+        public bool MmeDmaWriteMethodBarrierV => (MmeDmaWriteMethodBarrier & 0x1) != 0;
+        public fixed uint ReservedDF0[2436];
+        public MmeShadowScratch SetMmeShadowScratch;
+#pragma warning restore CS0649
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Twod/TwodTexture.cs b/src/Ryujinx.Graphics.Gpu/Engine/Twod/TwodTexture.cs
new file mode 100644
index 00000000..c28da094
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Twod/TwodTexture.cs
@@ -0,0 +1,22 @@
+using Ryujinx.Graphics.Gpu.Engine.Types;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Twod
+{
+    /// <summary>
+    /// Texture to texture (with optional resizing) copy parameters.
+    /// </summary>
+    struct TwodTexture
+    {
+#pragma warning disable CS0649
+        public ColorFormat Format;
+        public Boolean32 LinearLayout;
+        public MemoryLayout MemoryLayout;
+        public int Depth;
+        public int Layer;
+        public int Stride;
+        public int Width;
+        public int Height;
+        public GpuVa Address;
+#pragma warning restore CS0649
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Types/Boolean32.cs b/src/Ryujinx.Graphics.Gpu/Engine/Types/Boolean32.cs
new file mode 100644
index 00000000..c982347a
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Types/Boolean32.cs
@@ -0,0 +1,17 @@
+namespace Ryujinx.Graphics.Gpu.Engine.Types
+{
+    /// <summary>
+    /// Boolean value, stored as a 32-bits integer in memory.
+    /// </summary>
+    struct Boolean32
+    {
+#pragma warning disable CS0649
+        private uint _value;
+#pragma warning restore CS0649
+
+        public static implicit operator bool(Boolean32 value)
+        {
+            return (value._value & 1) != 0;
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Types/ColorFormat.cs b/src/Ryujinx.Graphics.Gpu/Engine/Types/ColorFormat.cs
new file mode 100644
index 00000000..889b5c8b
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Types/ColorFormat.cs
@@ -0,0 +1,165 @@
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Image;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Types
+{
+    /// <summary>
+    /// Color texture format.
+    /// </summary>
+    enum ColorFormat
+    {
+        R32G32B32A32Float = 0xc0,
+        R32G32B32A32Sint  = 0xc1,
+        R32G32B32A32Uint  = 0xc2,
+        R32G32B32X32Float = 0xc3,
+        R32G32B32X32Sint  = 0xc4,
+        R32G32B32X32Uint  = 0xc5,
+        R16G16B16X16Unorm = 0xc6,
+        R16G16B16X16Snorm = 0xc7,
+        R16G16B16X16Sint  = 0xc8,
+        R16G16B16X16Uint  = 0xc9,
+        R16G16B16A16Float = 0xca,
+        R32G32Float       = 0xcb,
+        R32G32Sint        = 0xcc,
+        R32G32Uint        = 0xcd,
+        R16G16B16X16Float = 0xce,
+        B8G8R8A8Unorm     = 0xcf,
+        B8G8R8A8Srgb      = 0xd0,
+        R10G10B10A2Unorm  = 0xd1,
+        R10G10B10A2Uint   = 0xd2,
+        R8G8B8A8Unorm     = 0xd5,
+        R8G8B8A8Srgb      = 0xd6,
+        R8G8B8X8Snorm     = 0xd7,
+        R8G8B8X8Sint      = 0xd8,
+        R8G8B8X8Uint      = 0xd9,
+        R16G16Unorm       = 0xda,
+        R16G16Snorm       = 0xdb,
+        R16G16Sint        = 0xdc,
+        R16G16Uint        = 0xdd,
+        R16G16Float       = 0xde,
+        R11G11B10Float    = 0xe0,
+        R32Sint           = 0xe3,
+        R32Uint           = 0xe4,
+        R32Float          = 0xe5,
+        B8G8R8X8Unorm     = 0xe6,
+        B8G8R8X8Srgb      = 0xe7,
+        B5G6R5Unorm       = 0xe8,
+        B5G5R5A1Unorm     = 0xe9,
+        R8G8Unorm         = 0xea,
+        R8G8Snorm         = 0xeb,
+        R8G8Sint          = 0xec,
+        R8G8Uint          = 0xed,
+        R16Unorm          = 0xee,
+        R16Snorm          = 0xef,
+        R16Sint           = 0xf0,
+        R16Uint           = 0xf1,
+        R16Float          = 0xf2,
+        R8Unorm           = 0xf3,
+        R8Snorm           = 0xf4,
+        R8Sint            = 0xf5,
+        R8Uint            = 0xf6,
+        B5G5R5X1Unorm     = 0xf8,
+        R8G8B8X8Unorm     = 0xf9,
+        R8G8B8X8Srgb      = 0xfa
+    }
+
+    static class ColorFormatConverter
+    {
+        /// <summary>
+        /// Converts the color texture format to a host compatible format.
+        /// </summary>
+        /// <param name="format">Color format</param>
+        /// <returns>Host compatible format enum value</returns>
+        public static FormatInfo Convert(this ColorFormat format)
+        {
+            return format switch
+            {
+                ColorFormat.R32G32B32A32Float => new FormatInfo(Format.R32G32B32A32Float, 1, 1, 16, 4),
+                ColorFormat.R32G32B32A32Sint  => new FormatInfo(Format.R32G32B32A32Sint,  1, 1, 16, 4),
+                ColorFormat.R32G32B32A32Uint  => new FormatInfo(Format.R32G32B32A32Uint,  1, 1, 16, 4),
+                ColorFormat.R32G32B32X32Float => new FormatInfo(Format.R32G32B32A32Float, 1, 1, 16, 4),
+                ColorFormat.R32G32B32X32Sint  => new FormatInfo(Format.R32G32B32A32Sint,  1, 1, 16, 4),
+                ColorFormat.R32G32B32X32Uint  => new FormatInfo(Format.R32G32B32A32Uint,  1, 1, 16, 4),
+                ColorFormat.R16G16B16X16Unorm => new FormatInfo(Format.R16G16B16A16Unorm, 1, 1, 8,  4),
+                ColorFormat.R16G16B16X16Snorm => new FormatInfo(Format.R16G16B16A16Snorm, 1, 1, 8,  4),
+                ColorFormat.R16G16B16X16Sint  => new FormatInfo(Format.R16G16B16A16Sint,  1, 1, 8,  4),
+                ColorFormat.R16G16B16X16Uint  => new FormatInfo(Format.R16G16B16A16Uint,  1, 1, 8,  4),
+                ColorFormat.R16G16B16A16Float => new FormatInfo(Format.R16G16B16A16Float, 1, 1, 8,  4),
+                ColorFormat.R32G32Float       => new FormatInfo(Format.R32G32Float,       1, 1, 8,  2),
+                ColorFormat.R32G32Sint        => new FormatInfo(Format.R32G32Sint,        1, 1, 8,  2),
+                ColorFormat.R32G32Uint        => new FormatInfo(Format.R32G32Uint,        1, 1, 8,  2),
+                ColorFormat.R16G16B16X16Float => new FormatInfo(Format.R16G16B16A16Float, 1, 1, 8,  4),
+                ColorFormat.B8G8R8A8Unorm     => new FormatInfo(Format.B8G8R8A8Unorm,     1, 1, 4,  4),
+                ColorFormat.B8G8R8A8Srgb      => new FormatInfo(Format.B8G8R8A8Srgb,      1, 1, 4,  4),
+                ColorFormat.R10G10B10A2Unorm  => new FormatInfo(Format.R10G10B10A2Unorm,  1, 1, 4,  4),
+                ColorFormat.R10G10B10A2Uint   => new FormatInfo(Format.R10G10B10A2Uint,   1, 1, 4,  4),
+                ColorFormat.R8G8B8A8Unorm     => new FormatInfo(Format.R8G8B8A8Unorm,     1, 1, 4,  4),
+                ColorFormat.R8G8B8A8Srgb      => new FormatInfo(Format.R8G8B8A8Srgb,      1, 1, 4,  4),
+                ColorFormat.R8G8B8X8Snorm     => new FormatInfo(Format.R8G8B8A8Snorm,     1, 1, 4,  4),
+                ColorFormat.R8G8B8X8Sint      => new FormatInfo(Format.R8G8B8A8Sint,      1, 1, 4,  4),
+                ColorFormat.R8G8B8X8Uint      => new FormatInfo(Format.R8G8B8A8Uint,      1, 1, 4,  4),
+                ColorFormat.R16G16Unorm       => new FormatInfo(Format.R16G16Unorm,       1, 1, 4,  2),
+                ColorFormat.R16G16Snorm       => new FormatInfo(Format.R16G16Snorm,       1, 1, 4,  2),
+                ColorFormat.R16G16Sint        => new FormatInfo(Format.R16G16Sint,        1, 1, 4,  2),
+                ColorFormat.R16G16Uint        => new FormatInfo(Format.R16G16Uint,        1, 1, 4,  2),
+                ColorFormat.R16G16Float       => new FormatInfo(Format.R16G16Float,       1, 1, 4,  2),
+                ColorFormat.R11G11B10Float    => new FormatInfo(Format.R11G11B10Float,    1, 1, 4,  3),
+                ColorFormat.R32Sint           => new FormatInfo(Format.R32Sint,           1, 1, 4,  1),
+                ColorFormat.R32Uint           => new FormatInfo(Format.R32Uint,           1, 1, 4,  1),
+                ColorFormat.R32Float          => new FormatInfo(Format.R32Float,          1, 1, 4,  1),
+                ColorFormat.B8G8R8X8Unorm     => new FormatInfo(Format.B8G8R8A8Unorm,     1, 1, 4,  4),
+                ColorFormat.B8G8R8X8Srgb      => new FormatInfo(Format.B8G8R8A8Srgb,      1, 1, 4,  4),
+                ColorFormat.B5G6R5Unorm       => new FormatInfo(Format.B5G6R5Unorm,       1, 1, 2,  3),
+                ColorFormat.B5G5R5A1Unorm     => new FormatInfo(Format.B5G5R5A1Unorm,     1, 1, 2,  4),
+                ColorFormat.R8G8Unorm         => new FormatInfo(Format.R8G8Unorm,         1, 1, 2,  2),
+                ColorFormat.R8G8Snorm         => new FormatInfo(Format.R8G8Snorm,         1, 1, 2,  2),
+                ColorFormat.R8G8Sint          => new FormatInfo(Format.R8G8Sint,          1, 1, 2,  2),
+                ColorFormat.R8G8Uint          => new FormatInfo(Format.R8G8Uint,          1, 1, 2,  2),
+                ColorFormat.R16Unorm          => new FormatInfo(Format.R16Unorm,          1, 1, 2,  1),
+                ColorFormat.R16Snorm          => new FormatInfo(Format.R16Snorm,          1, 1, 2,  1),
+                ColorFormat.R16Sint           => new FormatInfo(Format.R16Sint,           1, 1, 2,  1),
+                ColorFormat.R16Uint           => new FormatInfo(Format.R16Uint,           1, 1, 2,  1),
+                ColorFormat.R16Float          => new FormatInfo(Format.R16Float,          1, 1, 2,  1),
+                ColorFormat.R8Unorm           => new FormatInfo(Format.R8Unorm,           1, 1, 1,  1),
+                ColorFormat.R8Snorm           => new FormatInfo(Format.R8Snorm,           1, 1, 1,  1),
+                ColorFormat.R8Sint            => new FormatInfo(Format.R8Sint,            1, 1, 1,  1),
+                ColorFormat.R8Uint            => new FormatInfo(Format.R8Uint,            1, 1, 1,  1),
+                ColorFormat.B5G5R5X1Unorm     => new FormatInfo(Format.B5G5R5A1Unorm,     1, 1, 2,  4),
+                ColorFormat.R8G8B8X8Unorm     => new FormatInfo(Format.R8G8B8A8Unorm,     1, 1, 4,  4),
+                ColorFormat.R8G8B8X8Srgb      => new FormatInfo(Format.R8G8B8A8Srgb,      1, 1, 4,  4),
+                _                             => FormatInfo.Default
+            };
+        }
+
+        /// <summary>
+        /// Checks if a format has an alpha component.
+        /// </summary>
+        /// <param name="format">Format to be checked</param>
+        /// <returns>True if the format has no alpha component (RGBX), false if it does (RGBA)</returns>
+        public static bool NoAlpha(this ColorFormat format)
+        {
+            switch (format)
+            {
+                case ColorFormat.R32G32B32X32Float:
+                case ColorFormat.R32G32B32X32Sint:
+                case ColorFormat.R32G32B32X32Uint:
+                case ColorFormat.R16G16B16X16Unorm:
+                case ColorFormat.R16G16B16X16Snorm:
+                case ColorFormat.R16G16B16X16Sint:
+                case ColorFormat.R16G16B16X16Uint:
+                case ColorFormat.R16G16B16X16Float:
+                case ColorFormat.R8G8B8X8Snorm:
+                case ColorFormat.R8G8B8X8Sint:
+                case ColorFormat.R8G8B8X8Uint:
+                case ColorFormat.B8G8R8X8Unorm:
+                case ColorFormat.B8G8R8X8Srgb:
+                case ColorFormat.B5G5R5X1Unorm:
+                case ColorFormat.R8G8B8X8Unorm:
+                case ColorFormat.R8G8B8X8Srgb:
+                    return true;
+            }
+
+            return false;
+        }
+    }
+}
\ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Types/GpuVa.cs b/src/Ryujinx.Graphics.Gpu/Engine/Types/GpuVa.cs
new file mode 100644
index 00000000..839faac9
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Types/GpuVa.cs
@@ -0,0 +1,22 @@
+namespace Ryujinx.Graphics.Gpu.Engine.Types
+{
+    /// <summary>
+    /// Split GPU virtual address.
+    /// </summary>
+    struct GpuVa
+    {
+#pragma warning disable CS0649
+        public uint High;
+        public uint Low;
+#pragma warning restore CS0649
+
+        /// <summary>
+        /// Packs the split address into a 64-bits address value.
+        /// </summary>
+        /// <returns>The 64-bits address value</returns>
+        public ulong Pack()
+        {
+            return Low | ((ulong)High << 32);
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Types/MemoryLayout.cs b/src/Ryujinx.Graphics.Gpu/Engine/Types/MemoryLayout.cs
new file mode 100644
index 00000000..6da96bd4
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Types/MemoryLayout.cs
@@ -0,0 +1,37 @@
+namespace Ryujinx.Graphics.Gpu.Engine.Types
+{
+    /// <summary>
+    /// Memory layout parameters, for block linear textures.
+    /// </summary>
+    struct MemoryLayout
+    {
+#pragma warning disable CS0649
+        public uint Packed;
+#pragma warning restore CS0649
+
+        public int UnpackGobBlocksInX()
+        {
+            return 1 << (int)(Packed & 0xf);
+        }
+
+        public int UnpackGobBlocksInY()
+        {
+            return 1 << (int)((Packed >> 4) & 0xf);
+        }
+
+        public int UnpackGobBlocksInZ()
+        {
+            return 1 << (int)((Packed >> 8) & 0xf);
+        }
+
+        public bool UnpackIsLinear()
+        {
+            return (Packed & 0x1000) != 0;
+        }
+
+        public bool UnpackIsTarget3D()
+        {
+            return (Packed & 0x10000) != 0;
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Types/PrimitiveType.cs b/src/Ryujinx.Graphics.Gpu/Engine/Types/PrimitiveType.cs
new file mode 100644
index 00000000..dae63124
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Types/PrimitiveType.cs
@@ -0,0 +1,99 @@
+using Ryujinx.Graphics.GAL;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Types
+{
+    /// <summary>
+    /// Draw primitive type.
+    /// </summary>
+    enum PrimitiveType
+    {
+        Points,
+        Lines,
+        LineLoop,
+        LineStrip,
+        Triangles,
+        TriangleStrip,
+        TriangleFan,
+        Quads,
+        QuadStrip,
+        Polygon,
+        LinesAdjacency,
+        LineStripAdjacency,
+        TrianglesAdjacency,
+        TriangleStripAdjacency,
+        Patches
+    }
+
+    /// <summary>
+    /// Alternative primitive type that might override <see cref="PrimitiveType"/>.
+    /// </summary>
+    enum PrimitiveTypeOverride
+    {
+        Points = 1,
+        Lines = 2,
+        LineStrip = 3,
+        Triangles = 4,
+        TriangleStrip = 5,
+        TriangleFan = 0x1015,
+        LinesAdjacency = 10,
+        LineStripAdjacency = 11,
+        TrianglesAdjacency = 12,
+        TriangleStripAdjacency = 13,
+        Patches = 14
+    }
+
+    static class PrimitiveTypeConverter
+    {
+        /// <summary>
+        /// Converts the primitive type into something that can be used with the host API.
+        /// </summary>
+        /// <param name="type">The primitive type to convert</param>
+        /// <returns>A host compatible enum value</returns>
+        public static PrimitiveTopology Convert(this PrimitiveType type)
+        {
+            return type switch
+            {
+                PrimitiveType.Points                 => PrimitiveTopology.Points,
+                PrimitiveType.Lines                  => PrimitiveTopology.Lines,
+                PrimitiveType.LineLoop               => PrimitiveTopology.LineLoop,
+                PrimitiveType.LineStrip              => PrimitiveTopology.LineStrip,
+                PrimitiveType.Triangles              => PrimitiveTopology.Triangles,
+                PrimitiveType.TriangleStrip          => PrimitiveTopology.TriangleStrip,
+                PrimitiveType.TriangleFan            => PrimitiveTopology.TriangleFan,
+                PrimitiveType.Quads                  => PrimitiveTopology.Quads,
+                PrimitiveType.QuadStrip              => PrimitiveTopology.QuadStrip,
+                PrimitiveType.Polygon                => PrimitiveTopology.Polygon,
+                PrimitiveType.LinesAdjacency         => PrimitiveTopology.LinesAdjacency,
+                PrimitiveType.LineStripAdjacency     => PrimitiveTopology.LineStripAdjacency,
+                PrimitiveType.TrianglesAdjacency     => PrimitiveTopology.TrianglesAdjacency,
+                PrimitiveType.TriangleStripAdjacency => PrimitiveTopology.TriangleStripAdjacency,
+                PrimitiveType.Patches                => PrimitiveTopology.Patches,
+                _                                    => PrimitiveTopology.Triangles
+            };
+        }
+
+        /// <summary>
+        /// Converts the primitive type into something that can be used with the host API.
+        /// </summary>
+        /// <param name="type">The primitive type to convert</param>
+        /// <returns>A host compatible enum value</returns>
+        public static PrimitiveTopology Convert(this PrimitiveTypeOverride type)
+        {
+            return type switch
+            {
+                PrimitiveTypeOverride.Points                 => PrimitiveTopology.Points,
+                PrimitiveTypeOverride.Lines                  => PrimitiveTopology.Lines,
+                PrimitiveTypeOverride.LineStrip              => PrimitiveTopology.LineStrip,
+                PrimitiveTypeOverride.Triangles              => PrimitiveTopology.Triangles,
+                PrimitiveTypeOverride.TriangleStrip          => PrimitiveTopology.TriangleStrip,
+                PrimitiveTypeOverride.TriangleFan            => PrimitiveTopology.TriangleFan,
+                PrimitiveTypeOverride.LinesAdjacency         => PrimitiveTopology.LinesAdjacency,
+                PrimitiveTypeOverride.LineStripAdjacency     => PrimitiveTopology.LineStripAdjacency,
+                PrimitiveTypeOverride.TrianglesAdjacency     => PrimitiveTopology.TrianglesAdjacency,
+                PrimitiveTypeOverride.TriangleStripAdjacency => PrimitiveTopology.TriangleStripAdjacency,
+                PrimitiveTypeOverride.Patches                => PrimitiveTopology.Patches,
+                _                                            => PrimitiveTopology.Triangles
+            };
+        }
+    }
+}
\ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Types/SamplerIndex.cs b/src/Ryujinx.Graphics.Gpu/Engine/Types/SamplerIndex.cs
new file mode 100644
index 00000000..839a4d0a
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Types/SamplerIndex.cs
@@ -0,0 +1,11 @@
+namespace Ryujinx.Graphics.Gpu.Engine.Types
+{
+    /// <summary>
+    /// Sampler pool indexing mode.
+    /// </summary>
+    enum SamplerIndex
+    {
+        Independently  = 0,
+        ViaHeaderIndex = 1
+    }
+}
\ No newline at end of file
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Types/SbDescriptor.cs b/src/Ryujinx.Graphics.Gpu/Engine/Types/SbDescriptor.cs
new file mode 100644
index 00000000..c457dbf9
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Types/SbDescriptor.cs
@@ -0,0 +1,20 @@
+namespace Ryujinx.Graphics.Gpu.Engine.Types
+{
+    /// <summary>
+    /// Storage buffer address and size information.
+    /// </summary>
+    struct SbDescriptor
+    {
+#pragma warning disable CS0649
+        public uint AddressLow;
+        public uint AddressHigh;
+        public int Size;
+        public int Padding;
+#pragma warning restore CS0649
+
+        public ulong PackAddress()
+        {
+            return AddressLow | ((ulong)AddressHigh << 32);
+        }
+    }
+}
diff --git a/src/Ryujinx.Graphics.Gpu/Engine/Types/ZetaFormat.cs b/src/Ryujinx.Graphics.Gpu/Engine/Types/ZetaFormat.cs
new file mode 100644
index 00000000..1de1621f
--- /dev/null
+++ b/src/Ryujinx.Graphics.Gpu/Engine/Types/ZetaFormat.cs
@@ -0,0 +1,42 @@
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.Image;
+
+namespace Ryujinx.Graphics.Gpu.Engine.Types
+{
+    /// <summary>
+    /// Depth-stencil texture format.
+    /// </summary>
+    enum ZetaFormat
+    {
+        D32Float       = 0xa,
+        D16Unorm       = 0x13,
+        D24UnormS8Uint = 0x14,
+        D24Unorm       = 0x15,
+        S8UintD24Unorm = 0x16,
+        S8Uint         = 0x17,
+        D32FloatS8Uint = 0x19
+    }
+
+    static class ZetaFormatConverter
+    {
+        /// <summary>
+        /// Converts the depth-stencil texture format to a host compatible format.
+        /// </summary>
+        /// <param name="format">Depth-stencil format</param>
+        /// <returns>Host compatible format enum value</returns>
+        public static FormatInfo Convert(this ZetaFormat format)
+        {
+            return format switch
+            {
+                ZetaFormat.D32Float          => new FormatInfo(Format.D32Float,       1, 1, 4, 1),
+                ZetaFormat.D16Unorm          => new FormatInfo(Format.D16Unorm,       1, 1, 2, 1),
+                ZetaFormat.D24UnormS8Uint    => new FormatInfo(Format.D24UnormS8Uint, 1, 1, 4, 2),
+                ZetaFormat.D24Unorm          => new FormatInfo(Format.D24UnormS8Uint, 1, 1, 4, 1),
+                ZetaFormat.S8UintD24Unorm    => new FormatInfo(Format.S8UintD24Unorm, 1, 1, 4, 2),
+                ZetaFormat.S8Uint            => new FormatInfo(Format.S8Uint,         1, 1, 1, 1),
+                ZetaFormat.D32FloatS8Uint    => new FormatInfo(Format.D32FloatS8Uint, 1, 1, 8, 2),
+                _                            => FormatInfo.Default
+            };
+        }
+    }
+}