Initial work

17 files changed, 1991 insertions, 0 deletions

diff --git a/Ryujinx.Graphics.Gpu/Engine/Compute.cs b/Ryujinx.Graphics.Gpu/Engine/Compute.cs
new file mode 100644
index 00000000..c8627435
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/Compute.cs
@@ -0,0 +1,83 @@
+using Ryujinx.Graphics.Gpu.State;
+using Ryujinx.Graphics.Shader;
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    partial class Methods
+    {
+        public void Dispatch(int argument)
+        {
+            uint dispatchParamsAddress = (uint)_context.State.Get<int>(MethodOffset.DispatchParamsAddress);
+
+            var dispatchParams = _context.MemoryAccessor.Read<ComputeParams>((ulong)dispatchParamsAddress << 8);
+
+            GpuVa shaderBaseAddress = _context.State.Get<GpuVa>(MethodOffset.ShaderBaseAddress);
+
+            ulong shaderGpuVa = shaderBaseAddress.Pack() + (uint)dispatchParams.ShaderOffset;
+
+            ComputeShader cs = _shaderCache.GetComputeShader(
+                shaderGpuVa,
+                dispatchParams.UnpackBlockSizeX(),
+                dispatchParams.UnpackBlockSizeY(),
+                dispatchParams.UnpackBlockSizeZ());
+
+            _context.Renderer.ComputePipeline.SetProgram(cs.Interface);
+
+            ShaderProgramInfo info = cs.Shader.Info;
+
+            uint sbEnableMask = 0;
+            uint ubEnableMask = dispatchParams.UnpackUniformBuffersEnableMask();
+
+            for (int index = 0; index < dispatchParams.UniformBuffers.Length; index++)
+            {
+                if ((ubEnableMask & (1 << index)) == 0)
+                {
+                    continue;
+                }
+
+                ulong gpuVa = dispatchParams.UniformBuffers[index].PackAddress();
+                ulong size  = dispatchParams.UniformBuffers[index].UnpackSize();
+
+                _bufferManager.SetComputeUniformBuffer(index, gpuVa, size);
+            }
+
+            for (int index = 0; index < info.SBuffers.Count; index++)
+            {
+                BufferDescriptor sb = info.SBuffers[index];
+
+                sbEnableMask |= 1u << sb.Slot;
+
+                ulong sbDescAddress = _bufferManager.GetComputeUniformBufferAddress(0);
+
+                int sbDescOffset = 0x310 + sb.Slot * 0x10;
+
+                sbDescAddress += (ulong)sbDescOffset;
+
+                Span<byte> sbDescriptorData = _context.PhysicalMemory.Read(sbDescAddress, 0x10);
+
+                SbDescriptor sbDescriptor = MemoryMarshal.Cast<byte, SbDescriptor>(sbDescriptorData)[0];
+
+                _bufferManager.SetComputeStorageBuffer(sb.Slot, sbDescriptor.PackAddress(), (uint)sbDescriptor.Size);
+            }
+
+            ubEnableMask = 0;
+
+            for (int index = 0; index < info.CBuffers.Count; index++)
+            {
+                ubEnableMask |= 1u << info.CBuffers[index].Slot;
+            }
+
+            _bufferManager.SetComputeStorageBufferEnableMask(sbEnableMask);
+            _bufferManager.SetComputeUniformBufferEnableMask(ubEnableMask);
+
+            _bufferManager.CommitComputeBindings();
+
+            _context.Renderer.ComputePipeline.Dispatch(
+                dispatchParams.UnpackGridSizeX(),
+                dispatchParams.UnpackGridSizeY(),
+                dispatchParams.UnpackGridSizeZ());
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/ComputeParams.cs b/Ryujinx.Graphics.Gpu/Engine/ComputeParams.cs
new file mode 100644
index 00000000..03582f05
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/ComputeParams.cs
@@ -0,0 +1,126 @@
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    struct UniformBufferParams
+    {
+        public int AddressLow;
+        public int AddressHighAndSize;
+
+        public ulong PackAddress()
+        {
+            return (uint)AddressLow | ((ulong)(AddressHighAndSize & 0xff) << 32);
+        }
+
+        public ulong UnpackSize()
+        {
+            return (ulong)((AddressHighAndSize >> 15) & 0x1ffff);
+        }
+    }
+
+    struct ComputeParams
+    {
+        public int Unknown0;
+        public int Unknown1;
+        public int Unknown2;
+        public int Unknown3;
+        public int Unknown4;
+        public int Unknown5;
+        public int Unknown6;
+        public int Unknown7;
+        public int ShaderOffset;
+        public int Unknown9;
+        public int Unknown10;
+        public int Unknown11;
+        public int GridSizeX;
+        public int GridSizeYZ;
+        public int Unknown14;
+        public int Unknown15;
+        public int Unknown16;
+        public int Unknown17;
+        public int BlockSizeX;
+        public int BlockSizeYZ;
+        public int UniformBuffersConfig;
+        public int Unknown21;
+        public int Unknown22;
+        public int Unknown23;
+        public int Unknown24;
+        public int Unknown25;
+        public int Unknown26;
+        public int Unknown27;
+        public int Unknown28;
+
+        private UniformBufferParams _uniformBuffer0;
+        private UniformBufferParams _uniformBuffer1;
+        private UniformBufferParams _uniformBuffer2;
+        private UniformBufferParams _uniformBuffer3;
+        private UniformBufferParams _uniformBuffer4;
+        private UniformBufferParams _uniformBuffer5;
+        private UniformBufferParams _uniformBuffer6;
+        private UniformBufferParams _uniformBuffer7;
+
+        public Span<UniformBufferParams> UniformBuffers
+        {
+            get
+            {
+                return MemoryMarshal.CreateSpan(ref _uniformBuffer0, 8);
+            }
+        }
+
+        public int Unknown45;
+        public int Unknown46;
+        public int Unknown47;
+        public int Unknown48;
+        public int Unknown49;
+        public int Unknown50;
+        public int Unknown51;
+        public int Unknown52;
+        public int Unknown53;
+        public int Unknown54;
+        public int Unknown55;
+        public int Unknown56;
+        public int Unknown57;
+        public int Unknown58;
+        public int Unknown59;
+        public int Unknown60;
+        public int Unknown61;
+        public int Unknown62;
+        public int Unknown63;
+
+        public int UnpackGridSizeX()
+        {
+            return GridSizeX & 0x7fffffff;
+        }
+
+        public int UnpackGridSizeY()
+        {
+            return GridSizeYZ & 0xffff;
+        }
+
+        public int UnpackGridSizeZ()
+        {
+            return (GridSizeYZ >> 16) & 0xffff;
+        }
+
+        public int UnpackBlockSizeX()
+        {
+            return (BlockSizeX >> 16) & 0xffff;
+        }
+
+        public int UnpackBlockSizeY()
+        {
+            return BlockSizeYZ & 0xffff;
+        }
+
+        public int UnpackBlockSizeZ()
+        {
+            return (BlockSizeYZ >> 16) & 0xffff;
+        }
+
+        public uint UnpackUniformBuffersEnableMask()
+        {
+            return (uint)UniformBuffersConfig & 0xff;
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/ComputeShader.cs b/Ryujinx.Graphics.Gpu/Engine/ComputeShader.cs
new file mode 100644
index 00000000..cc7d4d99
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/ComputeShader.cs
@@ -0,0 +1,18 @@
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Shader;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    class ComputeShader
+    {
+        public IProgram Interface { get; set; }
+
+        public ShaderProgram Shader { get; }
+
+        public ComputeShader(IProgram program, ShaderProgram shader)
+        {
+            Interface = program;
+            Shader    = shader;
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/GraphicsShader.cs b/Ryujinx.Graphics.Gpu/Engine/GraphicsShader.cs
new file mode 100644
index 00000000..a8ccc05a
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/GraphicsShader.cs
@@ -0,0 +1,17 @@
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Shader;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    class GraphicsShader
+    {
+        public IProgram Interface { get; set; }
+
+        public ShaderProgram[] Shader { get; }
+
+        public GraphicsShader()
+        {
+            Shader = new ShaderProgram[5];
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/Inline2Memory.cs b/Ryujinx.Graphics.Gpu/Engine/Inline2Memory.cs
new file mode 100644
index 00000000..d2816ac5
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/Inline2Memory.cs
@@ -0,0 +1,42 @@
+using Ryujinx.Graphics.Gpu.State;
+using System;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    partial class Methods
+    {
+        private Inline2MemoryParams _params;
+
+        private bool _isLinear;
+
+        private int _offset;
+        private int _size;
+
+        public void Execute(int argument)
+        {
+            _params = _context.State.Get<Inline2MemoryParams>(MethodOffset.Inline2MemoryParams);
+
+            _isLinear = (argument & 1) != 0;
+
+            _offset = 0;
+            _size   = _params.LineLengthIn * _params.LineCount;
+        }
+
+        public void PushData(int argument)
+        {
+            if (_isLinear)
+            {
+                for (int shift = 0; shift < 32 && _offset < _size; shift += 8, _offset++)
+                {
+                    ulong gpuVa = _params.DstAddress.Pack() + (ulong)_offset;
+
+                    _context.MemoryAccessor.Write(gpuVa, new byte[] { (byte)(argument >> shift) });
+                }
+            }
+            else
+            {
+                throw new NotImplementedException();
+            }
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/MethodClear.cs b/Ryujinx.Graphics.Gpu/Engine/MethodClear.cs
new file mode 100644
index 00000000..b4680fa5
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/MethodClear.cs
@@ -0,0 +1,55 @@
+using Ryujinx.Graphics.GAL.Color;
+using Ryujinx.Graphics.Gpu.State;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    partial class Methods
+    {
+        private void Clear(int argument)
+        {
+            UpdateState();
+
+            bool clearDepth   = (argument & 1) != 0;
+            bool clearStencil = (argument & 2) != 0;
+
+            uint componentMask = (uint)((argument >> 2) & 0xf);
+
+            int index = (argument >> 6) & 0xf;
+
+            if (componentMask != 0)
+            {
+                ClearColors clearColor = _context.State.GetClearColors();
+
+                ColorF color = new ColorF(
+                    clearColor.Red,
+                    clearColor.Green,
+                    clearColor.Blue,
+                    clearColor.Alpha);
+
+                _context.Renderer.GraphicsPipeline.ClearRenderTargetColor(
+                    index,
+                    componentMask,
+                    color);
+            }
+
+            if (clearDepth || clearStencil)
+            {
+                float depthValue   = _context.State.GetClearDepthValue();
+                int   stencilValue = _context.State.GetClearStencilValue();
+
+                int stencilMask = 0;
+
+                if (clearStencil)
+                {
+                    stencilMask = _context.State.GetStencilTestState().FrontMask;
+                }
+
+                _context.Renderer.GraphicsPipeline.ClearRenderTargetDepthStencil(
+                    depthValue,
+                    clearDepth,
+                    stencilValue,
+                    stencilMask);
+            }
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/MethodCopyBuffer.cs b/Ryujinx.Graphics.Gpu/Engine/MethodCopyBuffer.cs
new file mode 100644
index 00000000..19ffb0e3
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/MethodCopyBuffer.cs
@@ -0,0 +1,79 @@
+using Ryujinx.Graphics.Gpu.State;
+using Ryujinx.Graphics.Texture;
+using System;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    partial class Methods
+    {
+        private void CopyBuffer(int argument)
+        {
+            var cbp = _context.State.Get<CopyBufferParams>(MethodOffset.CopyBufferParams);
+
+            var swizzle = _context.State.Get<CopyBufferSwizzle>(MethodOffset.CopyBufferSwizzle);
+
+            bool srcLinear = (argument & (1 << 7)) != 0;
+            bool dstLinear = (argument & (1 << 8)) != 0;
+            bool copy2D    = (argument & (1 << 9)) != 0;
+
+            int size = cbp.XCount;
+
+            if (size == 0)
+            {
+                return;
+            }
+
+            if (copy2D)
+            {
+                // Buffer to texture copy.
+                int srcBpp = swizzle.UnpackSrcComponentsCount() * swizzle.UnpackComponentSize();
+                int dstBpp = swizzle.UnpackDstComponentsCount() * swizzle.UnpackComponentSize();
+
+                var dst = _context.State.Get<CopyBufferTexture>(MethodOffset.CopyBufferDstTexture);
+                var src = _context.State.Get<CopyBufferTexture>(MethodOffset.CopyBufferSrcTexture);
+
+                var srcCalculator = new OffsetCalculator(
+                    src.Width,
+                    src.Height,
+                    cbp.SrcStride,
+                    srcLinear,
+                    src.MemoryLayout.UnpackGobBlocksInY(),
+                    srcBpp);
+
+                var dstCalculator = new OffsetCalculator(
+                    dst.Width,
+                    dst.Height,
+                    cbp.DstStride,
+                    dstLinear,
+                    dst.MemoryLayout.UnpackGobBlocksInY(),
+                    dstBpp);
+
+                ulong srcBaseAddress = _context.MemoryManager.Translate(cbp.SrcAddress.Pack());
+                ulong dstBaseAddress = _context.MemoryManager.Translate(cbp.DstAddress.Pack());
+
+                for (int y = 0; y < cbp.YCount; y++)
+                for (int x = 0; x < cbp.XCount; x++)
+                {
+                    int srcOffset = srcCalculator.GetOffset(src.RegionX + x, src.RegionY + y);
+                    int dstOffset = dstCalculator.GetOffset(dst.RegionX + x, dst.RegionY + y);
+
+                    ulong srcAddress = srcBaseAddress + (ulong)srcOffset;
+                    ulong dstAddress = dstBaseAddress + (ulong)dstOffset;
+
+                    Span<byte> pixel = _context.PhysicalMemory.Read(srcAddress, (ulong)srcBpp);
+
+                    _context.PhysicalMemory.Write(dstAddress, pixel);
+                }
+            }
+            else
+            {
+                // Buffer to buffer copy.
+                _bufferManager.CopyBuffer(cbp.SrcAddress, cbp.DstAddress, (uint)size);
+
+                Span<byte> data = _context.MemoryAccessor.Read(cbp.SrcAddress.Pack(), (uint)size);
+
+                _context.MemoryAccessor.Write(cbp.DstAddress.Pack(), data);
+            }
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/MethodCopyTexture.cs b/Ryujinx.Graphics.Gpu/Engine/MethodCopyTexture.cs
new file mode 100644
index 00000000..e2c40805
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/MethodCopyTexture.cs
@@ -0,0 +1,70 @@
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.State;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    partial class Methods
+    {
+        private void CopyTexture(int argument)
+        {
+            CopyTexture dstCopyTexture = _context.State.GetCopyDstTexture();
+            CopyTexture srcCopyTexture = _context.State.GetCopySrcTexture();
+
+            Image.Texture srcTexture = _textureManager.FindOrCreateTexture(srcCopyTexture);
+
+            if (srcTexture == null)
+            {
+                return;
+            }
+
+            // 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 == Format.D32Float)
+            {
+                dstCopyTexture.Format = RtFormat.D32Float;
+            }
+
+            Image.Texture dstTexture = _textureManager.FindOrCreateTexture(dstCopyTexture);
+
+            if (dstTexture == null)
+            {
+                return;
+            }
+
+            CopyTextureControl control = _context.State.GetCopyTextureControl();
+
+            CopyRegion region = _context.State.GetCopyRegion();
+
+            int srcX1 = (int)(region.SrcXF >> 32);
+            int srcY1 = (int)(region.SrcYF >> 32);
+
+            int srcX2 = (int)((region.SrcXF + region.SrcWidthRF  * region.DstWidth)  >> 32);
+            int srcY2 = (int)((region.SrcYF + region.SrcHeightRF * region.DstHeight) >> 32);
+
+            int dstX1 = region.DstX;
+            int dstY1 = region.DstY;
+
+            int dstX2 = region.DstX + region.DstWidth;
+            int dstY2 = region.DstY + region.DstHeight;
+
+            Extents2D srcRegion = new Extents2D(
+                srcX1 / srcTexture.Info.SamplesInX,
+                srcY1 / srcTexture.Info.SamplesInY,
+                srcX2 / srcTexture.Info.SamplesInX,
+                srcY2 / srcTexture.Info.SamplesInY);
+
+            Extents2D dstRegion = new Extents2D(
+                dstX1 / dstTexture.Info.SamplesInX,
+                dstY1 / dstTexture.Info.SamplesInY,
+                dstX2 / dstTexture.Info.SamplesInX,
+                dstY2 / dstTexture.Info.SamplesInY);
+
+            bool linearFilter = control.UnpackLinearFilter();
+
+            srcTexture.HostTexture.CopyTo(dstTexture.HostTexture, srcRegion, dstRegion, linearFilter);
+
+            dstTexture.Modified = true;
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/MethodDraw.cs b/Ryujinx.Graphics.Gpu/Engine/MethodDraw.cs
new file mode 100644
index 00000000..dd360113
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/MethodDraw.cs
@@ -0,0 +1,133 @@
+using Ryujinx.Graphics.Gpu.State;
+using Ryujinx.Graphics.Gpu.Image;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    partial class Methods
+    {
+        private bool _drawIndexed;
+
+        private int _firstIndex;
+        private int _indexCount;
+
+        private bool _instancedHasState;
+        private bool _instancedIndexed;
+
+        private int _instancedFirstIndex;
+        private int _instancedFirstVertex;
+        private int _instancedFirstInstance;
+        private int _instancedIndexCount;
+        private int _instancedDrawStateFirst;
+        private int _instancedDrawStateCount;
+
+        private int _instanceIndex;
+
+        public PrimitiveType PrimitiveType { get; private set; }
+
+        private void DrawEnd(int argument)
+        {
+            UpdateState();
+
+            bool instanced = _vsUsesInstanceId || _isAnyVbInstanced;
+
+            if (instanced)
+            {
+                if (!_instancedHasState)
+                {
+                    _instancedHasState = true;
+
+                    _instancedIndexed = _drawIndexed;
+
+                    _instancedFirstIndex    = _firstIndex;
+                    _instancedFirstVertex   = _context.State.GetBaseVertex();
+                    _instancedFirstInstance = _context.State.GetBaseInstance();
+
+                    _instancedIndexCount = _indexCount;
+
+                    VertexBufferDrawState drawState = _context.State.GetVertexBufferDrawState();
+
+                    _instancedDrawStateFirst = drawState.First;
+                    _instancedDrawStateCount = drawState.Count;
+                }
+
+                return;
+            }
+
+            int firstInstance = _context.State.GetBaseInstance();
+
+            if (_drawIndexed)
+            {
+                _drawIndexed = false;
+
+                int firstVertex = _context.State.GetBaseVertex();
+
+                _context.Renderer.GraphicsPipeline.DrawIndexed(
+                    _indexCount,
+                    1,
+                    _firstIndex,
+                    firstVertex,
+                    firstInstance);
+            }
+            else
+            {
+                VertexBufferDrawState drawState = _context.State.GetVertexBufferDrawState();
+
+                _context.Renderer.GraphicsPipeline.Draw(
+                    drawState.Count,
+                    1,
+                    drawState.First,
+                    firstInstance);
+            }
+        }
+
+        private void DrawBegin(int argument)
+        {
+            PrimitiveType type = (PrimitiveType)(argument & 0xffff);
+
+            _context.Renderer.GraphicsPipeline.SetPrimitiveTopology(type.Convert());
+
+            PrimitiveType = type;
+
+            if ((argument & (1 << 26)) != 0)
+            {
+                _instanceIndex++;
+            }
+            else if ((argument & (1 << 27)) == 0)
+            {
+                _instanceIndex = 0;
+            }
+        }
+
+        private void SetIndexCount(int argument)
+        {
+            _drawIndexed = true;
+        }
+
+        public void PerformDeferredDraws()
+        {
+            // Perform any pending instanced draw.
+            if (_instancedHasState)
+            {
+                _instancedHasState = false;
+
+                if (_instancedIndexed)
+                {
+                    _context.Renderer.GraphicsPipeline.DrawIndexed(
+                        _instancedIndexCount,
+                        _instanceIndex + 1,
+                        _instancedFirstIndex,
+                        _instancedFirstVertex,
+                        _instancedFirstInstance);
+                }
+                else
+                {
+                    _context.Renderer.GraphicsPipeline.Draw(
+                        _instancedDrawStateCount,
+                        _instanceIndex + 1,
+                        _instancedDrawStateFirst,
+                        _instancedFirstInstance);
+                }
+            }
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/MethodReport.cs b/Ryujinx.Graphics.Gpu/Engine/MethodReport.cs
new file mode 100644
index 00000000..fd0a31a1
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/MethodReport.cs
@@ -0,0 +1,100 @@
+using Ryujinx.Common;
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.State;
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    partial class Methods
+    {
+        private ulong _runningCounter;
+
+        private void Report(int argument)
+        {
+            ReportMode mode = (ReportMode)(argument & 3);
+
+            ReportCounterType type = (ReportCounterType)((argument >> 23) & 0x1f);
+
+            switch (mode)
+            {
+                case ReportMode.Semaphore: ReportSemaphore();   break;
+                case ReportMode.Counter:   ReportCounter(type); break;
+            }
+        }
+
+        private void ReportSemaphore()
+        {
+            ReportState state = _context.State.GetReportState();
+
+            _context.MemoryAccessor.Write(state.Address.Pack(), state.Payload);
+
+            _context.AdvanceSequence();
+        }
+
+        private struct CounterData
+        {
+            public ulong Counter;
+            public ulong Timestamp;
+        }
+
+        private void ReportCounter(ReportCounterType type)
+        {
+            CounterData counterData = new CounterData();
+
+            ulong counter = 0;
+
+            switch (type)
+            {
+                case ReportCounterType.Zero:
+                    counter = 0;
+                    break;
+                case ReportCounterType.SamplesPassed:
+                    counter = _context.Renderer.GetCounter(CounterType.SamplesPassed);
+                    break;
+                case ReportCounterType.PrimitivesGenerated:
+                    counter = _context.Renderer.GetCounter(CounterType.PrimitivesGenerated);
+                    break;
+                case ReportCounterType.TransformFeedbackPrimitivesWritten:
+                    counter = _context.Renderer.GetCounter(CounterType.TransformFeedbackPrimitivesWritten);
+                    break;
+            }
+
+            ulong ticks;
+
+            if (GraphicsConfig.FastGpuTime)
+            {
+                ticks = _runningCounter++;
+            }
+            else
+            {
+                ticks = ConvertNanosecondsToTicks((ulong)PerformanceCounter.ElapsedNanoseconds);
+            }
+
+            counterData.Counter   = counter;
+            counterData.Timestamp = ticks;
+
+            Span<CounterData> counterDataSpan = MemoryMarshal.CreateSpan(ref counterData, 1);
+
+            Span<byte> data = MemoryMarshal.Cast<CounterData, byte>(counterDataSpan);
+
+            ReportState state = _context.State.GetReportState();
+
+            _context.MemoryAccessor.Write(state.Address.Pack(), data);
+        }
+
+        private static ulong ConvertNanosecondsToTicks(ulong nanoseconds)
+        {
+            // We need to divide first to avoid overflows.
+            // We fix up the result later by calculating the difference and adding
+            // that to the result.
+            ulong divided = nanoseconds / 625;
+
+            ulong rounded = divided * 625;
+
+            ulong errorBias = ((nanoseconds - rounded) * 384) / 625;
+
+            return divided * 384 + errorBias;
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/MethodResetCounter.cs b/Ryujinx.Graphics.Gpu/Engine/MethodResetCounter.cs
new file mode 100644
index 00000000..12a87845
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/MethodResetCounter.cs
@@ -0,0 +1,26 @@
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.State;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    partial class Methods
+    {
+        private 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;
+            }
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/MethodUniformBufferBind.cs b/Ryujinx.Graphics.Gpu/Engine/MethodUniformBufferBind.cs
new file mode 100644
index 00000000..a9ebce83
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/MethodUniformBufferBind.cs
@@ -0,0 +1,52 @@
+using Ryujinx.Graphics.Gpu.State;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    partial class Methods
+    {
+        private void UniformBufferBind0(int argument)
+        {
+            UniformBufferBind(argument, ShaderType.Vertex);
+        }
+
+        private void UniformBufferBind1(int argument)
+        {
+            UniformBufferBind(argument, ShaderType.TessellationControl);
+        }
+
+        private void UniformBufferBind2(int argument)
+        {
+            UniformBufferBind(argument, ShaderType.TessellationEvaluation);
+        }
+
+        private void UniformBufferBind3(int argument)
+        {
+            UniformBufferBind(argument, ShaderType.Geometry);
+        }
+
+        private void UniformBufferBind4(int argument)
+        {
+            UniformBufferBind(argument, ShaderType.Fragment);
+        }
+
+        private void UniformBufferBind(int argument, ShaderType type)
+        {
+            bool enable = (argument & 1) != 0;
+
+            int index = (argument >> 4) & 0x1f;
+
+            if (enable)
+            {
+                UniformBufferState uniformBuffer = _context.State.GetUniformBufferState();
+
+                ulong address = uniformBuffer.Address.Pack();
+
+                _bufferManager.SetGraphicsUniformBuffer((int)type, index, address, (uint)uniformBuffer.Size);
+            }
+            else
+            {
+                _bufferManager.SetGraphicsUniformBuffer((int)type, index, 0, 0);
+            }
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/MethodUniformBufferUpdate.cs b/Ryujinx.Graphics.Gpu/Engine/MethodUniformBufferUpdate.cs
new file mode 100644
index 00000000..f0b8e2d7
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/MethodUniformBufferUpdate.cs
@@ -0,0 +1,18 @@
+using Ryujinx.Graphics.Gpu.State;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    partial class Methods
+    {
+        private void UniformBufferUpdate(int argument)
+        {
+            UniformBufferState uniformBuffer = _context.State.GetUniformBufferState();
+
+            _context.MemoryAccessor.Write(uniformBuffer.Address.Pack() + (uint)uniformBuffer.Offset, argument);
+
+            _context.State.SetUniformBufferOffset(uniformBuffer.Offset + 4);
+
+            _context.AdvanceSequence();
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/Methods.cs b/Ryujinx.Graphics.Gpu/Engine/Methods.cs
new file mode 100644
index 00000000..db72a861
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/Methods.cs
@@ -0,0 +1,784 @@
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.GAL.Blend;
+using Ryujinx.Graphics.GAL.DepthStencil;
+using Ryujinx.Graphics.GAL.InputAssembler;
+using Ryujinx.Graphics.GAL.Texture;
+using Ryujinx.Graphics.Gpu.Image;
+using Ryujinx.Graphics.Gpu.Memory;
+using Ryujinx.Graphics.Gpu.State;
+using Ryujinx.Graphics.Shader;
+using System;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    partial class Methods
+    {
+        private GpuContext _context;
+
+        private ShaderCache _shaderCache;
+
+        private BufferManager  _bufferManager;
+        private TextureManager _textureManager;
+
+        public TextureManager TextureManager => _textureManager;
+
+        private bool _isAnyVbInstanced;
+        private bool _vsUsesInstanceId;
+
+        public Methods(GpuContext context)
+        {
+            _context = context;
+
+            _shaderCache = new ShaderCache(_context);
+
+            _bufferManager  = new BufferManager(context);
+            _textureManager = new TextureManager(context, _bufferManager);
+
+            RegisterCallbacks();
+        }
+
+        private void RegisterCallbacks()
+        {
+            _context.State.RegisterCopyBufferCallback(CopyBuffer);
+            _context.State.RegisterCopyTextureCallback(CopyTexture);
+
+            _context.State.RegisterDrawEndCallback(DrawEnd);
+
+            _context.State.RegisterDrawBeginCallback(DrawBegin);
+
+            _context.State.RegisterSetIndexCountCallback(SetIndexCount);
+
+            _context.State.RegisterClearCallback(Clear);
+
+            _context.State.RegisterReportCallback(Report);
+
+            _context.State.RegisterUniformBufferUpdateCallback(UniformBufferUpdate);
+
+            _context.State.RegisterUniformBufferBind0Callback(UniformBufferBind0);
+            _context.State.RegisterUniformBufferBind1Callback(UniformBufferBind1);
+            _context.State.RegisterUniformBufferBind2Callback(UniformBufferBind2);
+            _context.State.RegisterUniformBufferBind3Callback(UniformBufferBind3);
+            _context.State.RegisterUniformBufferBind4Callback(UniformBufferBind4);
+
+            _context.State.RegisterCallback(MethodOffset.InvalidateTextures, InvalidateTextures);
+
+            _context.State.RegisterCallback(MethodOffset.ResetCounter, ResetCounter);
+
+            _context.State.RegisterCallback(MethodOffset.Inline2MemoryExecute,  Execute);
+            _context.State.RegisterCallback(MethodOffset.Inline2MemoryPushData, PushData);
+
+            _context.State.RegisterCallback(MethodOffset.Dispatch, Dispatch);
+        }
+
+        public Image.Texture GetTexture(ulong address) => _textureManager.Find2(address);
+
+        private void UpdateState()
+        {
+            if ((_context.State.StateWriteFlags & StateWriteFlags.Any) == 0)
+            {
+                CommitBindings();
+
+                return;
+            }
+
+            // Shaders must be the first one to be updated if modified, because
+            // some of the other state depends on information from the currently
+            // bound shaders.
+            if ((_context.State.StateWriteFlags & StateWriteFlags.ShaderState) != 0)
+            {
+                UpdateShaderState();
+            }
+
+            if ((_context.State.StateWriteFlags & StateWriteFlags.RenderTargetGroup) != 0)
+            {
+                UpdateRenderTargetGroupState();
+            }
+
+            if ((_context.State.StateWriteFlags & StateWriteFlags.DepthTestState) != 0)
+            {
+                UpdateDepthTestState();
+            }
+
+            if ((_context.State.StateWriteFlags & StateWriteFlags.ViewportTransform) != 0)
+            {
+                UpdateViewportTransform();
+            }
+
+            if ((_context.State.StateWriteFlags & StateWriteFlags.DepthBiasState) != 0)
+            {
+                UpdateDepthBiasState();
+            }
+
+            if ((_context.State.StateWriteFlags & StateWriteFlags.StencilTestState) != 0)
+            {
+                UpdateStencilTestState();
+            }
+
+            if ((_context.State.StateWriteFlags & StateWriteFlags.SamplerPoolState) != 0)
+            {
+                UpdateSamplerPoolState();
+            }
+
+            if ((_context.State.StateWriteFlags & StateWriteFlags.TexturePoolState) != 0)
+            {
+                UpdateTexturePoolState();
+            }
+
+            if ((_context.State.StateWriteFlags & StateWriteFlags.InputAssemblerGroup) != 0)
+            {
+                UpdateInputAssemblerGroupState();
+            }
+
+            if ((_context.State.StateWriteFlags & StateWriteFlags.FaceState) != 0)
+            {
+                UpdateFaceState();
+            }
+
+            if ((_context.State.StateWriteFlags & StateWriteFlags.RtColorMask) != 0)
+            {
+                UpdateRtColorMask();
+            }
+
+            if ((_context.State.StateWriteFlags & StateWriteFlags.BlendState) != 0)
+            {
+                UpdateBlendState();
+            }
+
+            _context.State.StateWriteFlags &= ~StateWriteFlags.Any;
+
+            CommitBindings();
+        }
+
+        private void CommitBindings()
+        {
+            _bufferManager.CommitBindings();
+            _textureManager.CommitBindings();
+        }
+
+        public void InvalidateRange(ulong address, ulong size)
+        {
+            _bufferManager.InvalidateRange(address, size);
+            _textureManager.InvalidateRange(address, size);
+        }
+
+        public void InvalidateTextureRange(ulong address, ulong size)
+        {
+            _textureManager.InvalidateRange(address, size);
+        }
+
+        private void UpdateRenderTargetGroupState()
+        {
+            TextureMsaaMode msaaMode = _context.State.GetRtMsaaMode();
+
+            int samplesInX = msaaMode.SamplesInX();
+            int samplesInY = msaaMode.SamplesInY();
+
+            Image.Texture color3D = Get3DRenderTarget(samplesInX, samplesInY);
+
+            if (color3D == null)
+            {
+                for (int index = 0; index < Constants.TotalRenderTargets; index++)
+                {
+                    RtColorState colorState = _context.State.GetRtColorState(index);
+
+                    if (!IsRtEnabled(colorState))
+                    {
+                        _textureManager.SetRenderTargetColor(index, null);
+
+                        continue;
+                    }
+
+                    Image.Texture color = _textureManager.FindOrCreateTexture(
+                        colorState,
+                        samplesInX,
+                        samplesInY);
+
+                    _textureManager.SetRenderTargetColor(index, color);
+
+                    color.Modified = true;
+                }
+            }
+            else
+            {
+                _textureManager.SetRenderTargetColor3D(color3D);
+
+                color3D.Modified = true;
+            }
+
+            bool dsEnable = _context.State.Get<bool>(MethodOffset.RtDepthStencilEnable);
+
+            Image.Texture depthStencil = null;
+
+            if (dsEnable)
+            {
+                var dsState = _context.State.GetRtDepthStencilState();
+                var dsSize  = _context.State.GetRtDepthStencilSize();
+
+                depthStencil = _textureManager.FindOrCreateTexture(
+                    dsState,
+                    dsSize,
+                    samplesInX,
+                    samplesInY);
+            }
+
+            _textureManager.SetRenderTargetDepthStencil(depthStencil);
+        }
+
+        private Image.Texture Get3DRenderTarget(int samplesInX, int samplesInY)
+        {
+            RtColorState colorState0 = _context.State.GetRtColorState(0);
+
+            if (!IsRtEnabled(colorState0) || !colorState0.MemoryLayout.UnpackIsTarget3D() || colorState0.Depth != 1)
+            {
+                return null;
+            }
+
+            int slices = 1;
+            int unused = 0;
+
+            for (int index = 1; index < Constants.TotalRenderTargets; index++)
+            {
+                RtColorState colorState = _context.State.GetRtColorState(index);
+
+                if (!IsRtEnabled(colorState))
+                {
+                    unused++;
+
+                    continue;
+                }
+
+                if (colorState.MemoryLayout.UnpackIsTarget3D() && colorState.Depth == 1)
+                {
+                    slices++;
+                }
+            }
+
+            if (slices + unused == Constants.TotalRenderTargets)
+            {
+                colorState0.Depth = slices;
+
+                return _textureManager.FindOrCreateTexture(colorState0, samplesInX, samplesInY);
+            }
+
+            return null;
+        }
+
+        private static bool IsRtEnabled(RtColorState colorState)
+        {
+            // Colors are disabled by writing 0 to the format.
+            return colorState.Format != 0 && colorState.WidthOrStride != 0;
+        }
+
+        private void UpdateDepthTestState()
+        {
+            _context.Renderer.GraphicsPipeline.SetDepthTest(new DepthTestDescriptor(
+                _context.State.GetDepthTestEnable().IsTrue(),
+                _context.State.GetDepthWriteEnable().IsTrue(),
+                _context.State.GetDepthTestFunc()));
+        }
+
+        private void UpdateViewportTransform()
+        {
+            Viewport[] viewports = new Viewport[Constants.TotalViewports];
+
+            for (int index = 0; index < Constants.TotalViewports; index++)
+            {
+                var transform = _context.State.Get<ViewportTransform>(MethodOffset.ViewportTransform + index * 8);
+                var extents   = _context.State.Get<ViewportExtents>  (MethodOffset.ViewportExtents   + index * 4);
+
+                float x = transform.TranslateX - MathF.Abs(transform.ScaleX);
+                float y = transform.TranslateY - MathF.Abs(transform.ScaleY);
+
+                float width  = transform.ScaleX * 2;
+                float height = transform.ScaleY * 2;
+
+                RectangleF region = new RectangleF(x, y, width, height);
+
+                viewports[index] = new Viewport(
+                    region,
+                    transform.UnpackSwizzleX(),
+                    transform.UnpackSwizzleY(),
+                    transform.UnpackSwizzleZ(),
+                    transform.UnpackSwizzleW(),
+                    extents.DepthNear,
+                    extents.DepthFar);
+            }
+
+            _context.Renderer.GraphicsPipeline.SetViewports(0, viewports);
+        }
+
+        private void UpdateDepthBiasState()
+        {
+            var polygonOffset = _context.State.Get<DepthBiasState>(MethodOffset.DepthBiasState);
+
+            float factor = _context.State.Get<float>(MethodOffset.DepthBiasFactor);
+            float units  = _context.State.Get<float>(MethodOffset.DepthBiasUnits);
+            float clamp  = _context.State.Get<float>(MethodOffset.DepthBiasClamp);
+
+            PolygonModeMask enables = 0;
+
+            enables  = (polygonOffset.PointEnable.IsTrue() ? PolygonModeMask.Point : 0);
+            enables |= (polygonOffset.LineEnable.IsTrue()  ? PolygonModeMask.Line  : 0);
+            enables |= (polygonOffset.FillEnable.IsTrue()  ? PolygonModeMask.Fill  : 0);
+
+            _context.Renderer.GraphicsPipeline.SetDepthBias(enables, factor, units, clamp);
+        }
+
+        private void UpdateStencilTestState()
+        {
+            StencilBackMasks     backMasks = _context.State.GetStencilBackMasks();
+            StencilTestState     test      = _context.State.GetStencilTestState();
+            StencilBackTestState backTest  = _context.State.GetStencilBackTestState();
+
+            CompareOp backFunc;
+            StencilOp backSFail;
+            StencilOp backDpPass;
+            StencilOp backDpFail;
+            int       backFuncRef;
+            int       backFuncMask;
+            int       backMask;
+
+            if (backTest.TwoSided.IsTrue())
+            {
+                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;
+            }
+
+            _context.Renderer.GraphicsPipeline.SetStencilTest(new StencilTestDescriptor(
+                test.Enable.IsTrue(),
+                test.FrontFunc,
+                test.FrontSFail,
+                test.FrontDpPass,
+                test.FrontDpFail,
+                test.FrontFuncRef,
+                test.FrontFuncMask,
+                test.FrontMask,
+                backFunc,
+                backSFail,
+                backDpPass,
+                backDpFail,
+                backFuncRef,
+                backFuncMask,
+                backMask));
+        }
+
+        private void UpdateSamplerPoolState()
+        {
+            PoolState samplerPool = _context.State.GetSamplerPoolState();
+
+            _textureManager.SetSamplerPool(samplerPool.Address.Pack(), samplerPool.MaximumId);
+        }
+
+        private void UpdateTexturePoolState()
+        {
+            PoolState texturePool = _context.State.GetTexturePoolState();
+
+            _textureManager.SetTexturePool(texturePool.Address.Pack(), texturePool.MaximumId);
+
+            _textureManager.SetTextureBufferIndex(_context.State.GetTextureBufferIndex());
+        }
+
+        private void UpdateInputAssemblerGroupState()
+        {
+            // Must be updated before the vertex buffer.
+            if ((_context.State.StateWriteFlags & StateWriteFlags.VertexAttribState) != 0)
+            {
+                UpdateVertexAttribState();
+            }
+
+            if ((_context.State.StateWriteFlags & StateWriteFlags.PrimitiveRestartState) != 0)
+            {
+                UpdatePrimitiveRestartState();
+            }
+
+            if ((_context.State.StateWriteFlags & StateWriteFlags.IndexBufferState) != 0)
+            {
+                UpdateIndexBufferState();
+            }
+
+            if ((_context.State.StateWriteFlags & StateWriteFlags.VertexBufferState) != 0)
+            {
+                UpdateVertexBufferState();
+            }
+        }
+
+        private void UpdateVertexAttribState()
+        {
+            VertexAttribDescriptor[] vertexAttribs = new VertexAttribDescriptor[16];
+
+            for (int index = 0; index < 16; index++)
+            {
+                VertexAttribState vertexAttrib = _context.State.GetVertexAttribState(index);
+
+                if (!FormatTable.TryGetAttribFormat(vertexAttrib.UnpackFormat(), out Format format))
+                {
+                    // TODO: warning.
+
+                    format = Format.R32G32B32A32Float;
+                }
+
+                vertexAttribs[index] = new VertexAttribDescriptor(
+                    vertexAttrib.UnpackBufferIndex(),
+                    vertexAttrib.UnpackOffset(),
+                    format);
+            }
+
+            _context.Renderer.GraphicsPipeline.BindVertexAttribs(vertexAttribs);
+        }
+
+        private void UpdatePrimitiveRestartState()
+        {
+            PrimitiveRestartState primitiveRestart = _context.State.Get<PrimitiveRestartState>(MethodOffset.PrimitiveRestartState);
+
+            _context.Renderer.GraphicsPipeline.SetPrimitiveRestart(
+                primitiveRestart.Enable,
+                primitiveRestart.Index);
+        }
+
+        private void UpdateIndexBufferState()
+        {
+            IndexBufferState indexBuffer = _context.State.GetIndexBufferState();
+
+            _firstIndex = indexBuffer.First;
+            _indexCount = indexBuffer.Count;
+
+            if (_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)(_firstIndex + _indexCount);
+
+            switch (indexBuffer.Type)
+            {
+                case IndexType.UShort: size *= 2; break;
+                case IndexType.UInt:   size *= 4; break;
+            }
+
+            _bufferManager.SetIndexBuffer(gpuVa, size, indexBuffer.Type);
+
+            // The index buffer affects the vertex buffer size calculation, we
+            // need to ensure that they are updated.
+            UpdateVertexBufferState();
+        }
+
+        private uint GetIndexBufferMaxIndex(ulong gpuVa, ulong size, IndexType type)
+        {
+            ulong address = _context.MemoryManager.Translate(gpuVa);
+
+            Span<byte> data = _context.PhysicalMemory.Read(address, size);
+
+            uint maxIndex = 0;
+
+            switch (type)
+            {
+                case IndexType.UByte:
+                {
+                    for (int index = 0; index < data.Length; index++)
+                    {
+                        if (maxIndex < data[index])
+                        {
+                            maxIndex = data[index];
+                        }
+                    }
+
+                    break;
+                }
+
+                case IndexType.UShort:
+                {
+                    Span<ushort> indices = MemoryMarshal.Cast<byte, ushort>(data);
+
+                    for (int index = 0; index < indices.Length; index++)
+                    {
+                        if (maxIndex < indices[index])
+                        {
+                            maxIndex = indices[index];
+                        }
+                    }
+
+                    break;
+                }
+
+                case IndexType.UInt:
+                {
+                    Span<uint> indices = MemoryMarshal.Cast<byte, uint>(data);
+
+                    for (int index = 0; index < indices.Length; index++)
+                    {
+                        if (maxIndex < indices[index])
+                        {
+                            maxIndex = indices[index];
+                        }
+                    }
+
+                    break;
+                }
+            }
+
+            return maxIndex;
+        }
+
+        private void UpdateVertexBufferState()
+        {
+            _isAnyVbInstanced = false;
+
+            for (int index = 0; index < 16; index++)
+            {
+                VertexBufferState vertexBuffer = _context.State.GetVertexBufferState(index);
+
+                if (!vertexBuffer.UnpackEnable())
+                {
+                    _bufferManager.SetVertexBuffer(index, 0, 0, 0, 0);
+
+                    continue;
+                }
+
+                GpuVa endAddress = _context.State.GetVertexBufferEndAddress(index);
+
+                ulong address = vertexBuffer.Address.Pack();
+
+                int stride = vertexBuffer.UnpackStride();
+
+                bool instanced = _context.State.Get<bool>(MethodOffset.VertexBufferInstanced + index);
+
+                int divisor = instanced ? vertexBuffer.Divisor : 0;
+
+                _isAnyVbInstanced |= divisor != 0;
+
+                ulong size;
+
+                if (_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 = endAddress.Pack() - address + 1;
+                }
+                else
+                {
+                    // For non-indexed draws, we can guess the size from the vertex count
+                    // and stride.
+                    int firstInstance = _context.State.GetBaseInstance();
+
+                    VertexBufferDrawState drawState = _context.State.GetVertexBufferDrawState();
+
+                    size = (ulong)((firstInstance + drawState.First + drawState.Count) * stride);
+                }
+
+                _bufferManager.SetVertexBuffer(index, address, size, stride, divisor);
+            }
+        }
+
+        private void UpdateFaceState()
+        {
+            FaceState face = _context.State.GetFaceState();
+
+            _context.Renderer.GraphicsPipeline.SetFaceCulling(face.CullEnable.IsTrue(), face.CullFace);
+
+            _context.Renderer.GraphicsPipeline.SetFrontFace(face.FrontFace);
+        }
+
+        private void UpdateRtColorMask()
+        {
+            uint[] componentMasks = new uint[Constants.TotalRenderTargets];
+
+            for (int index = 0; index < Constants.TotalRenderTargets; index++)
+            {
+                RtColorMask colorMask = _context.State.Get<RtColorMask>(MethodOffset.RtColorMask + index);
+
+                uint componentMask = 0;
+
+                componentMask  = (colorMask.UnpackRed()   ? 1u : 0u);
+                componentMask |= (colorMask.UnpackGreen() ? 2u : 0u);
+                componentMask |= (colorMask.UnpackBlue()  ? 4u : 0u);
+                componentMask |= (colorMask.UnpackAlpha() ? 8u : 0u);
+
+                componentMasks[index] = componentMask;
+            }
+
+            _context.Renderer.GraphicsPipeline.SetRenderTargetColorMasks(componentMasks);
+        }
+
+        private void UpdateBlendState()
+        {
+            BlendState[] blends = new BlendState[8];
+
+            for (int index = 0; index < 8; index++)
+            {
+                bool blendEnable = _context.State.GetBlendEnable(index).IsTrue();
+
+                BlendState blend = _context.State.GetBlendState(index);
+
+                BlendDescriptor descriptor = new BlendDescriptor(
+                    blendEnable,
+                    blend.ColorOp,
+                    blend.ColorSrcFactor,
+                    blend.ColorDstFactor,
+                    blend.AlphaOp,
+                    blend.AlphaSrcFactor,
+                    blend.AlphaDstFactor);
+
+                _context.Renderer.GraphicsPipeline.BindBlendState(index, descriptor);
+            }
+        }
+
+        private struct SbDescriptor
+        {
+            public uint AddressLow;
+            public uint AddressHigh;
+            public int  Size;
+            public int  Padding;
+
+            public ulong PackAddress()
+            {
+                return AddressLow | ((ulong)AddressHigh << 32);
+            }
+        }
+
+        private void UpdateShaderState()
+        {
+            ShaderAddresses addresses = new ShaderAddresses();
+
+            Span<ShaderAddresses> addressesSpan = MemoryMarshal.CreateSpan(ref addresses, 1);
+
+            Span<ulong> addressesArray = MemoryMarshal.Cast<ShaderAddresses, ulong>(addressesSpan);
+
+            ulong baseAddress = _context.State.GetShaderBaseAddress().Pack();
+
+            for (int index = 0; index < 6; index++)
+            {
+                ShaderState shader = _context.State.GetShaderState(index);
+
+                if (!shader.UnpackEnable() && index != 1)
+                {
+                    continue;
+                }
+
+                addressesArray[index] = baseAddress + shader.Offset;
+            }
+
+            GraphicsShader gs = _shaderCache.GetGraphicsShader(addresses);
+
+            _vsUsesInstanceId = gs.Shader[0].Info.UsesInstanceId;
+
+            for (int stage = 0; stage < Constants.TotalShaderStages; stage++)
+            {
+                ShaderProgramInfo info = gs.Shader[stage]?.Info;
+
+                if (info == null)
+                {
+                    continue;
+                }
+
+                var textureBindings = new TextureBindingInfo[info.Textures.Count];
+
+                for (int index = 0; index < info.Textures.Count; index++)
+                {
+                    var descriptor = info.Textures[index];
+
+                    Target target = GetTarget(descriptor.Target);
+
+                    textureBindings[index] = new TextureBindingInfo(target, descriptor.HandleIndex);
+                }
+
+                _textureManager.BindTextures(stage, textureBindings);
+
+                uint sbEnableMask = 0;
+                uint ubEnableMask = 0;
+
+                for (int index = 0; index < info.SBuffers.Count; index++)
+                {
+                    BufferDescriptor sb = info.SBuffers[index];
+
+                    sbEnableMask |= 1u << sb.Slot;
+
+                    ulong sbDescAddress = _bufferManager.GetGraphicsUniformBufferAddress(stage, 0);
+
+                    int sbDescOffset = 0x110 + stage * 0x100 + sb.Slot * 0x10;
+
+                    sbDescAddress += (ulong)sbDescOffset;
+
+                    Span<byte> sbDescriptorData = _context.PhysicalMemory.Read(sbDescAddress, 0x10);
+
+                    SbDescriptor sbDescriptor = MemoryMarshal.Cast<byte, SbDescriptor>(sbDescriptorData)[0];
+
+                    _bufferManager.SetGraphicsStorageBuffer(stage, sb.Slot, sbDescriptor.PackAddress(), (uint)sbDescriptor.Size);
+                }
+
+                for (int index = 0; index < info.CBuffers.Count; index++)
+                {
+                    ubEnableMask |= 1u << info.CBuffers[index].Slot;
+                }
+
+                _bufferManager.SetGraphicsStorageBufferEnableMask(stage, sbEnableMask);
+                _bufferManager.SetGraphicsUniformBufferEnableMask(stage, ubEnableMask);
+            }
+
+            _context.Renderer.GraphicsPipeline.BindProgram(gs.Interface);
+        }
+
+        private static Target GetTarget(Shader.TextureTarget target)
+        {
+            target &= ~Shader.TextureTarget.Shadow;
+
+            switch (target)
+            {
+                case Shader.TextureTarget.Texture1D:
+                    return Target.Texture1D;
+
+                case Shader.TextureTarget.Texture1D | Shader.TextureTarget.Array:
+                    return Target.Texture1DArray;
+
+                case Shader.TextureTarget.Texture2D:
+                    return Target.Texture2D;
+
+                case Shader.TextureTarget.Texture2D | Shader.TextureTarget.Array:
+                    return Target.Texture2DArray;
+
+                case Shader.TextureTarget.Texture2D | Shader.TextureTarget.Multisample:
+                    return Target.Texture2DMultisample;
+
+                case Shader.TextureTarget.Texture2D | Shader.TextureTarget.Multisample | Shader.TextureTarget.Array:
+                    return Target.Texture2DMultisampleArray;
+
+                case Shader.TextureTarget.Texture3D:
+                    return Target.Texture3D;
+
+                case Shader.TextureTarget.TextureCube:
+                    return Target.Cubemap;
+
+                case Shader.TextureTarget.TextureCube | Shader.TextureTarget.Array:
+                    return Target.CubemapArray;
+            }
+
+            // TODO: Warning.
+
+            return Target.Texture2D;
+        }
+
+        private void InvalidateTextures(int argument)
+        {
+            _textureManager.Flush();
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/ShaderAddresses.cs b/Ryujinx.Graphics.Gpu/Engine/ShaderAddresses.cs
new file mode 100644
index 00000000..368b5a17
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/ShaderAddresses.cs
@@ -0,0 +1,34 @@
+using System;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    struct ShaderAddresses : IEquatable<ShaderAddresses>
+    {
+        public ulong VertexA;
+        public ulong Vertex;
+        public ulong TessControl;
+        public ulong TessEvaluation;
+        public ulong Geometry;
+        public ulong Fragment;
+
+        public override bool Equals(object other)
+        {
+            return other is ShaderAddresses addresses && Equals(addresses);
+        }
+
+        public bool Equals(ShaderAddresses other)
+        {
+            return VertexA        == other.VertexA &&
+                   Vertex         == other.Vertex &&
+                   TessControl    == other.TessControl &&
+                   TessEvaluation == other.TessEvaluation &&
+                   Geometry       == other.Geometry &&
+                   Fragment       == other.Fragment;
+        }
+
+        public override int GetHashCode()
+        {
+            return HashCode.Combine(VertexA, Vertex, TessControl, TessEvaluation, Geometry, Fragment);
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/ShaderCache.cs b/Ryujinx.Graphics.Gpu/Engine/ShaderCache.cs
new file mode 100644
index 00000000..79a84a6d
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/ShaderCache.cs
@@ -0,0 +1,228 @@
+using Ryujinx.Graphics.GAL;
+using Ryujinx.Graphics.Gpu.State;
+using Ryujinx.Graphics.Shader;
+using Ryujinx.Graphics.Shader.Translation;
+using System;
+using System.Collections.Generic;
+using System.Globalization;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    class ShaderCache
+    {
+        private const int MaxProgramSize = 0x100000;
+
+        private GpuContext _context;
+
+        private ShaderDumper _dumper;
+
+        private Dictionary<ulong, ComputeShader> _cpPrograms;
+
+        private Dictionary<ShaderAddresses, GraphicsShader> _gpPrograms;
+
+        public ShaderCache(GpuContext context)
+        {
+            _context = context;
+
+            _dumper = new ShaderDumper(context);
+
+            _cpPrograms = new Dictionary<ulong, ComputeShader>();
+
+            _gpPrograms = new Dictionary<ShaderAddresses, GraphicsShader>();
+        }
+
+        public ComputeShader GetComputeShader(ulong gpuVa, int localSizeX, int localSizeY, int localSizeZ)
+        {
+            if (!_cpPrograms.TryGetValue(gpuVa, out ComputeShader cpShader))
+            {
+                ShaderProgram shader = TranslateComputeShader(gpuVa);
+
+                shader.Replace(DefineNames.LocalSizeX, localSizeX.ToString(CultureInfo.InvariantCulture));
+                shader.Replace(DefineNames.LocalSizeY, localSizeY.ToString(CultureInfo.InvariantCulture));
+                shader.Replace(DefineNames.LocalSizeZ, localSizeZ.ToString(CultureInfo.InvariantCulture));
+
+                IShader hostShader = _context.Renderer.CompileShader(shader);
+
+                IProgram program = _context.Renderer.CreateProgram(new IShader[] { hostShader });
+
+                cpShader = new ComputeShader(program, shader);
+
+                _cpPrograms.Add(gpuVa, cpShader);
+            }
+
+            return cpShader;
+        }
+
+        public GraphicsShader GetGraphicsShader(ShaderAddresses addresses)
+        {
+            if (!_gpPrograms.TryGetValue(addresses, out GraphicsShader gpShader))
+            {
+                gpShader = new GraphicsShader();
+
+                if (addresses.VertexA != 0)
+                {
+                    gpShader.Shader[0] = TranslateGraphicsShader(addresses.Vertex, addresses.VertexA);
+                }
+                else
+                {
+                    gpShader.Shader[0] = TranslateGraphicsShader(addresses.Vertex);
+                }
+
+                gpShader.Shader[1] = TranslateGraphicsShader(addresses.TessControl);
+                gpShader.Shader[2] = TranslateGraphicsShader(addresses.TessEvaluation);
+                gpShader.Shader[3] = TranslateGraphicsShader(addresses.Geometry);
+                gpShader.Shader[4] = TranslateGraphicsShader(addresses.Fragment);
+
+                BackpropQualifiers(gpShader);
+
+                List<IShader> shaders = new List<IShader>();
+
+                for (int stage = 0; stage < gpShader.Shader.Length; stage++)
+                {
+                    if (gpShader.Shader[stage] == null)
+                    {
+                        continue;
+                    }
+
+                    IShader shader = _context.Renderer.CompileShader(gpShader.Shader[stage]);
+
+                    shaders.Add(shader);
+                }
+
+                gpShader.Interface = _context.Renderer.CreateProgram(shaders.ToArray());
+
+                _gpPrograms.Add(addresses, gpShader);
+            }
+
+            return gpShader;
+        }
+
+        private ShaderProgram TranslateComputeShader(ulong gpuVa)
+        {
+            if (gpuVa == 0)
+            {
+                return null;
+            }
+
+            ShaderProgram program;
+
+            const TranslationFlags flags =
+                TranslationFlags.Compute |
+                TranslationFlags.Unspecialized;
+
+            TranslationConfig translationConfig = new TranslationConfig(0x10000, _dumper.CurrentDumpIndex, flags);
+
+            Span<byte> code = _context.MemoryAccessor.Read(gpuVa, MaxProgramSize);
+
+            program = Translator.Translate(code, translationConfig);
+
+            _dumper.Dump(gpuVa, compute : true);
+
+            return program;
+        }
+
+        private ShaderProgram TranslateGraphicsShader(ulong gpuVa, ulong gpuVaA = 0)
+        {
+            if (gpuVa == 0)
+            {
+                return null;
+            }
+
+            ShaderProgram program;
+
+            const TranslationFlags flags =
+                TranslationFlags.DebugMode |
+                TranslationFlags.Unspecialized;
+
+            TranslationConfig translationConfig = new TranslationConfig(0x10000, _dumper.CurrentDumpIndex, flags);
+
+            if (gpuVaA != 0)
+            {
+                Span<byte> codeA = _context.MemoryAccessor.Read(gpuVaA, MaxProgramSize);
+                Span<byte> codeB = _context.MemoryAccessor.Read(gpuVa,  MaxProgramSize);
+
+                program = Translator.Translate(codeA, codeB, translationConfig);
+
+                _dumper.Dump(gpuVaA, compute: false);
+                _dumper.Dump(gpuVa,  compute: false);
+            }
+            else
+            {
+                Span<byte> code = _context.MemoryAccessor.Read(gpuVa, MaxProgramSize);
+
+                program = Translator.Translate(code, translationConfig);
+
+                _dumper.Dump(gpuVa, compute: false);
+            }
+
+            if (program.Stage == ShaderStage.Geometry)
+            {
+                PrimitiveType primitiveType = _context.Methods.PrimitiveType;
+
+                string inPrimitive = "points";
+
+                switch (primitiveType)
+                {
+                    case PrimitiveType.Points:
+                        inPrimitive = "points";
+                        break;
+                    case PrimitiveType.Lines:
+                    case PrimitiveType.LineLoop:
+                    case PrimitiveType.LineStrip:
+                        inPrimitive = "lines";
+                        break;
+                    case PrimitiveType.LinesAdjacency:
+                    case PrimitiveType.LineStripAdjacency:
+                        inPrimitive = "lines_adjacency";
+                        break;
+                    case PrimitiveType.Triangles:
+                    case PrimitiveType.TriangleStrip:
+                    case PrimitiveType.TriangleFan:
+                        inPrimitive = "triangles";
+                        break;
+                    case PrimitiveType.TrianglesAdjacency:
+                    case PrimitiveType.TriangleStripAdjacency:
+                        inPrimitive = "triangles_adjacency";
+                        break;
+                }
+
+                program.Replace(DefineNames.InputTopologyName, inPrimitive);
+            }
+
+            return program;
+        }
+
+        private void BackpropQualifiers(GraphicsShader program)
+        {
+            ShaderProgram fragmentShader = program.Shader[4];
+
+            bool isFirst = true;
+
+            for (int stage = 3; stage >= 0; stage--)
+            {
+                if (program.Shader[stage] == null)
+                {
+                    continue;
+                }
+
+                // We need to iterate backwards, since we do name replacement,
+                // and it would otherwise replace a subset of the longer names.
+                for (int attr = 31; attr >= 0; attr--)
+                {
+                    string iq = fragmentShader?.Info.InterpolationQualifiers[attr].ToGlslQualifier() ?? string.Empty;
+
+                    if (isFirst && iq != string.Empty)
+                    {
+                        program.Shader[stage].Replace($"{DefineNames.OutQualifierPrefixName}{attr}", iq);
+                    }
+                    else
+                    {
+                        program.Shader[stage].Replace($"{DefineNames.OutQualifierPrefixName}{attr} ", string.Empty);
+                    }
+                }
+
+                isFirst = false;
+            }
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Gpu/Engine/ShaderDumper.cs b/Ryujinx.Graphics.Gpu/Engine/ShaderDumper.cs
new file mode 100644
index 00000000..fdcf0612
--- /dev/null
+++ b/Ryujinx.Graphics.Gpu/Engine/ShaderDumper.cs
@@ -0,0 +1,126 @@
+using System.IO;
+
+namespace Ryujinx.Graphics.Gpu.Engine
+{
+    class ShaderDumper
+    {
+        private const int ShaderHeaderSize = 0x50;
+
+        private GpuContext _context;
+
+        private string _runtimeDir;
+        private string _dumpPath;
+        private int    _dumpIndex;
+
+        public int CurrentDumpIndex => _dumpIndex;
+
+        public ShaderDumper(GpuContext context)
+        {
+            _context = context;
+
+            _dumpIndex = 1;
+        }
+
+        public void Dump(ulong gpuVa, bool compute)
+        {
+            _dumpPath = GraphicsConfig.ShadersDumpPath;
+
+            if (string.IsNullOrWhiteSpace(_dumpPath))
+            {
+                return;
+            }
+
+            string fileName = "Shader" + _dumpIndex.ToString("d4") + ".bin";
+
+            string fullPath = Path.Combine(FullDir(), fileName);
+            string codePath = Path.Combine(CodeDir(), fileName);
+
+            _dumpIndex++;
+
+            ulong headerSize = compute ? 0UL : ShaderHeaderSize;
+
+            using (FileStream fullFile = File.Create(fullPath))
+            using (FileStream codeFile = File.Create(codePath))
+            {
+                BinaryWriter fullWriter = new BinaryWriter(fullFile);
+                BinaryWriter codeWriter = new BinaryWriter(codeFile);
+
+                for (ulong i = 0; i < headerSize; i += 4)
+                {
+                    fullWriter.Write(_context.MemoryAccessor.ReadInt32(gpuVa + i));
+                }
+
+                ulong offset = 0;
+
+                ulong instruction = 0;
+
+                // Dump until a NOP instruction is found.
+                while ((instruction >> 48 & 0xfff8) != 0x50b0)
+                {
+                    uint word0 = (uint)_context.MemoryAccessor.ReadInt32(gpuVa + headerSize + offset + 0);
+                    uint word1 = (uint)_context.MemoryAccessor.ReadInt32(gpuVa + headerSize + offset + 4);
+
+                    instruction = word0 | (ulong)word1 << 32;
+
+                    // Zero instructions (other kind of NOP) stop immediately,
+                    // this is to avoid two rows of zeroes.
+                    if (instruction == 0)
+                    {
+                        break;
+                    }
+
+                    fullWriter.Write(instruction);
+                    codeWriter.Write(instruction);
+
+                    offset += 8;
+                }
+
+                // Align to meet nvdisasm requirements.
+                while (offset % 0x20 != 0)
+                {
+                    fullWriter.Write(0);
+                    codeWriter.Write(0);
+
+                    offset += 4;
+                }
+            }
+        }
+
+        private string FullDir()
+        {
+            return CreateAndReturn(Path.Combine(DumpDir(), "Full"));
+        }
+
+        private string CodeDir()
+        {
+            return CreateAndReturn(Path.Combine(DumpDir(), "Code"));
+        }
+
+        private string DumpDir()
+        {
+            if (string.IsNullOrEmpty(_runtimeDir))
+            {
+                int index = 1;
+
+                do
+                {
+                    _runtimeDir = Path.Combine(_dumpPath, "Dumps" + index.ToString("d2"));
+
+                    index++;
+                }
+                while (Directory.Exists(_runtimeDir));
+
+                Directory.CreateDirectory(_runtimeDir);
+            }
+
+            return _runtimeDir;
+        }
+
+        private static string CreateAndReturn(string dir)
+        {
+            Directory.CreateDirectory(dir);
+
+            return dir;
+        }
+    }
+}
\ No newline at end of file