piwalker/Ryujinx
1
0
Fork 0
mirror of https://github.com/Ryujinx-NX/Ryujinx.git synced 2024-11-14 13:07:41 -07:00
Code Issues Actions Packages Projects Releases Wiki Activity

Support texture rectangle targets (non-normalized coords)

Browse Source
This commit is contained in:
gdkchan 2019-12-16 01:59:46 -03:00 committed by Thog
parent 2eccc7023a
commit 9d7a142a48
23 changed files with 473 additions and 356 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Ryujinx.Graphics.Gpu/Engine/Compute.cs
View File

@ -21,7 +21,7 @@ namespace Ryujinx.Graphics.Gpu.Engine
ulong shaderGpuVa = shaderBaseAddress.Pack() + (uint)dispatchParams.ShaderOffset;
// Note: A size of 0 is also invalid, the size must be at least 1.
int sharedMemorySize = Math.Clamp(dispatchParams.SharedMemorySize & 0xffff, 4, _context.Capabilities.MaximumComputeSharedMemorySize);
int sharedMemorySize = Math.Clamp(dispatchParams.SharedMemorySize & 0xffff, 1, _context.Capabilities.MaximumComputeSharedMemorySize);
ComputeShader cs = _shaderCache.GetComputeShader(
shaderGpuVa,

6
Ryujinx.Graphics.Gpu/Engine/Methods.cs
View File

@ -671,9 +671,7 @@ namespace Ryujinx.Graphics.Gpu.Engine
addressesArray[index] = baseAddress + shader.Offset;
}
bool viewportTransformEnable = GetViewportTransformEnable(state);
GraphicsShader gs = _shaderCache.GetGraphicsShader(addresses, !viewportTransformEnable);
GraphicsShader gs = _shaderCache.GetGraphicsShader(state, addresses);
_vsUsesInstanceId = gs.Shader[0].Program.Info.UsesInstanceId;
@ -734,7 +732,7 @@ namespace Ryujinx.Graphics.Gpu.Engine
_context.Renderer.Pipeline.BindProgram(gs.HostProgram);
}
private bool GetViewportTransformEnable(GpuState state)
public bool GetViewportTransformEnable(GpuState state)
{
// FIXME: We should read ViewportTransformEnable, but it seems that some games writes 0 there?
// return state.Get<Boolean32>(MethodOffset.ViewportTransformEnable) != 0;

15
Ryujinx.Graphics.Gpu/Image/TextureBindingsManager.cs
View File

@ -199,6 +199,21 @@ namespace Ryujinx.Graphics.Gpu.Image
}
}
public TextureDescriptor GetTextureDescriptor(GpuState state, int stageIndex, int handle)
{
int packedId = ReadPackedId(stageIndex, handle);
int textureId = UnpackTextureId(packedId);
var poolState = state.Get<PoolState>(MethodOffset.TexturePoolState);
ulong poolAddress = _context.MemoryManager.Translate(poolState.Address.Pack());
TexturePool texturePool = _texturePoolCache.FindOrCreate(poolAddress, poolState.MaximumId);
return texturePool.GetDescriptor(textureId);
}
private int ReadPackedId(int stage, int wordOffset)
{
ulong address;

5
Ryujinx.Graphics.Gpu/Image/TextureDescriptor.cs
View File

@ -101,6 +101,11 @@ namespace Ryujinx.Graphics.Gpu.Image
return (int)((Word5 >> 16) & 0x3fff) + 1;
}
public bool UnpackTextureCoordNormalized()
{
return (Word5 & (1 << 31)) != 0;
}
public int UnpackBaseLevel()
{
return (int)(Word7 & 0xf);

5
Ryujinx.Graphics.Gpu/Image/TextureManager.cs
View File

@ -129,6 +129,11 @@ namespace Ryujinx.Graphics.Gpu.Image
UpdateRenderTargets();
}
public TextureDescriptor GetGraphicsTextureDescriptor(GpuState state, int stageIndex, int handle)
{
return _gpBindingsManager.GetTextureDescriptor(state, stageIndex, handle);
}
private void UpdateRenderTargets()
{
bool anyChanged = false;

15
Ryujinx.Graphics.Gpu/Image/TexturePool.cs
View File

@ -36,11 +36,7 @@ namespace Ryujinx.Graphics.Gpu.Image
if (texture == null)
{
ulong address = Address + (ulong)(uint)id * DescriptorSize;
Span<byte> data = Context.PhysicalMemory.Read(address, DescriptorSize);
TextureDescriptor descriptor = MemoryMarshal.Cast<byte, TextureDescriptor>(data)[0];
TextureDescriptor descriptor = GetDescriptor(id);
TextureInfo info = GetInfo(descriptor);
@ -66,6 +62,15 @@ namespace Ryujinx.Graphics.Gpu.Image
return texture;
}
public TextureDescriptor GetDescriptor(int id)
{
ulong address = Address + (ulong)(uint)id * DescriptorSize;
Span<byte> data = Context.PhysicalMemory.Read(address, DescriptorSize);
return MemoryMarshal.Cast<byte, TextureDescriptor>(data)[0];
}
protected override void InvalidateRangeImpl(ulong address, ulong size)
{
ulong endAddress = address + size;

4
Ryujinx.Graphics.Gpu/Image/TextureTarget.cs
View File

@ -11,7 +11,7 @@ namespace Ryujinx.Graphics.Gpu.Image
Texture1DArray,
Texture2DArray,
TextureBuffer,
Texture2DLinear,
Texture2DRect,
CubemapArray
}
@ -33,7 +33,7 @@ namespace Ryujinx.Graphics.Gpu.Image
{
case TextureTarget.Texture1D: return Target.Texture1D;
case TextureTarget.Texture2D: return Target.Texture2D;
case TextureTarget.Texture2DLinear: return Target.Texture2D;
case TextureTarget.Texture2DRect: return Target.Texture2D;
case TextureTarget.Texture3D: return Target.Texture3D;
case TextureTarget.Texture1DArray: return Target.Texture1DArray;
case TextureTarget.Texture2DArray: return Target.Texture2DArray;

181
Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs
View File

@ -1,10 +1,10 @@
using Ryujinx.Graphics.GAL;
using Ryujinx.Graphics.Gpu.Image;
using Ryujinx.Graphics.Gpu.State;
using Ryujinx.Graphics.Shader;
using Ryujinx.Graphics.Shader.Translation;
using System;
using System.Collections.Generic;
using System.Globalization;
using System.Runtime.InteropServices;
namespace Ryujinx.Graphics.Gpu.Shader
@ -13,6 +13,8 @@ namespace Ryujinx.Graphics.Gpu.Shader
{
private const int MaxProgramSize = 0x100000;
private const TranslationFlags DefaultFlags = TranslationFlags.DebugMode;
private GpuContext _context;
private ShaderDumper _dumper;
@ -69,7 +71,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
return cpShader;
}
public GraphicsShader GetGraphicsShader(ShaderAddresses addresses, bool dividePosXY)
public GraphicsShader GetGraphicsShader(GpuState state, ShaderAddresses addresses)
{
bool isCached = _gpPrograms.TryGetValue(addresses, out List<GraphicsShader> list);
@ -86,28 +88,19 @@ namespace Ryujinx.Graphics.Gpu.Shader
GraphicsShader gpShaders = new GraphicsShader();
TranslationFlags flags =
TranslationFlags.DebugMode |
TranslationFlags.Unspecialized;
if (dividePosXY)
{
flags |= TranslationFlags.DividePosXY;
}
if (addresses.VertexA != 0)
{
gpShaders.Shader[0] = TranslateGraphicsShader(flags, addresses.Vertex, addresses.VertexA);
gpShaders.Shader[0] = TranslateGraphicsShader(state, ShaderStage.Vertex, addresses.Vertex, addresses.VertexA);
}
else
{
gpShaders.Shader[0] = TranslateGraphicsShader(flags, addresses.Vertex);
gpShaders.Shader[0] = TranslateGraphicsShader(state, ShaderStage.Vertex, addresses.Vertex);
}
gpShaders.Shader[1] = TranslateGraphicsShader(flags, addresses.TessControl);
gpShaders.Shader[2] = TranslateGraphicsShader(flags, addresses.TessEvaluation);
gpShaders.Shader[3] = TranslateGraphicsShader(flags, addresses.Geometry);
gpShaders.Shader[4] = TranslateGraphicsShader(flags, addresses.Fragment);
gpShaders.Shader[1] = TranslateGraphicsShader(state, ShaderStage.TessellationControl, addresses.TessControl);
gpShaders.Shader[2] = TranslateGraphicsShader(state, ShaderStage.TessellationEvaluation, addresses.TessEvaluation);
gpShaders.Shader[3] = TranslateGraphicsShader(state, ShaderStage.Geometry, addresses.Geometry);
gpShaders.Shader[4] = TranslateGraphicsShader(state, ShaderStage.Fragment, addresses.Fragment);
BackpropQualifiers(gpShaders);
@ -199,25 +192,31 @@ namespace Ryujinx.Graphics.Gpu.Shader
return null;
}
ShaderProgram program;
QueryInfoCallback queryInfo = (QueryInfoName info, int index) =>
{
switch (info)
{
case QueryInfoName.ComputeLocalSizeX:
return localSizeX;
case QueryInfoName.ComputeLocalSizeY:
return localSizeY;
case QueryInfoName.ComputeLocalSizeZ:
return localSizeZ;
case QueryInfoName.ComputeSharedMemorySize:
return sharedMemorySize;
}
const TranslationFlags flags =
TranslationFlags.Compute |
TranslationFlags.DebugMode |
TranslationFlags.Unspecialized;
return QueryInfoCommon(info);
};
ShaderProgram program;
Span<byte> code = _context.MemoryAccessor.Read(gpuVa, MaxProgramSize);
program = Translator.Translate(code, GetShaderCapabilities(), flags);
program = Translator.Translate(code, queryInfo, DefaultFlags | TranslationFlags.Compute);
int[] codeCached = MemoryMarshal.Cast<byte, int>(code.Slice(0, program.Size)).ToArray();
program.Replace(DefineNames.SharedMemorySize, (sharedMemorySize / 4).ToString(CultureInfo.InvariantCulture));
program.Replace(DefineNames.LocalSizeX, localSizeX.ToString(CultureInfo.InvariantCulture));
program.Replace(DefineNames.LocalSizeY, localSizeY.ToString(CultureInfo.InvariantCulture));
program.Replace(DefineNames.LocalSizeZ, localSizeZ.ToString(CultureInfo.InvariantCulture));
_dumper.Dump(code, compute: true, out string fullPath, out string codePath);
if (fullPath != null && codePath != null)
@ -229,13 +228,30 @@ namespace Ryujinx.Graphics.Gpu.Shader
return new CachedShader(program, codeCached);
}
private CachedShader TranslateGraphicsShader(TranslationFlags flags, ulong gpuVa, ulong gpuVaA = 0)
private CachedShader TranslateGraphicsShader(GpuState state, ShaderStage stage, ulong gpuVa, ulong gpuVaA = 0)
{
if (gpuVa == 0)
{
return new CachedShader(null, null);
}
QueryInfoCallback queryInfo = (QueryInfoName info, int index) =>
{
switch (info)
{
case QueryInfoName.IsTextureBuffer:
return Convert.ToInt32(QueryIsTextureBuffer(state, (int)stage - 1, index));
case QueryInfoName.IsTextureRectangle:
return Convert.ToInt32(QueryIsTextureRectangle(state, (int)stage - 1, index));
case QueryInfoName.PrimitiveTopology:
return (int)GetPrimitiveTopology();
case QueryInfoName.ViewportTransformEnable:
return Convert.ToInt32(_context.Methods.GetViewportTransformEnable(state));
}
return QueryInfoCommon(info);
};
ShaderProgram program;
int[] codeCached = null;
@ -245,9 +261,9 @@ namespace Ryujinx.Graphics.Gpu.Shader
Span<byte> codeA = _context.MemoryAccessor.Read(gpuVaA, MaxProgramSize);
Span<byte> codeB = _context.MemoryAccessor.Read(gpuVa, MaxProgramSize);
program = Translator.Translate(codeA, codeB, GetShaderCapabilities(), flags);
program = Translator.Translate(codeA, codeB, queryInfo, DefaultFlags);
// TODO: We should also check "codeA" into account.
// TODO: We should also take "codeA" into account.
codeCached = MemoryMarshal.Cast<byte, int>(codeB.Slice(0, program.Size)).ToArray();
_dumper.Dump(codeA, compute: false, out string fullPathA, out string codePathA);
@ -265,7 +281,7 @@ namespace Ryujinx.Graphics.Gpu.Shader
{
Span<byte> code = _context.MemoryAccessor.Read(gpuVa, MaxProgramSize);
program = Translator.Translate(code, GetShaderCapabilities(), flags);
program = Translator.Translate(code, queryInfo, DefaultFlags);
codeCached = MemoryMarshal.Cast<byte, int>(code.Slice(0, program.Size)).ToArray();
@ -278,40 +294,6 @@ namespace Ryujinx.Graphics.Gpu.Shader
}
}
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);
}
ulong address = _context.MemoryManager.Translate(gpuVa);
return new CachedShader(program, codeCached);
@ -350,13 +332,66 @@ namespace Ryujinx.Graphics.Gpu.Shader
}
}
private ShaderCapabilities GetShaderCapabilities()
private InputTopology GetPrimitiveTopology()
{
return new ShaderCapabilities(
_context.Capabilities.MaximumViewportDimensions,
_context.Capabilities.MaximumComputeSharedMemorySize,
_context.Capabilities.StorageBufferOffsetAlignment,
_context.Capabilities.SupportsNonConstantTextureOffset);
switch (_context.Methods.PrimitiveType)
{
case PrimitiveType.Points:
return InputTopology.Points;
case PrimitiveType.Lines:
case PrimitiveType.LineLoop:
case PrimitiveType.LineStrip:
return InputTopology.Lines;
case PrimitiveType.LinesAdjacency:
case PrimitiveType.LineStripAdjacency:
return InputTopology.LinesAdjacency;
case PrimitiveType.Triangles:
case PrimitiveType.TriangleStrip:
case PrimitiveType.TriangleFan:
return InputTopology.Triangles;
case PrimitiveType.TrianglesAdjacency:
case PrimitiveType.TriangleStripAdjacency:
return InputTopology.TrianglesAdjacency;
}
return InputTopology.Points;
}
private bool QueryIsTextureBuffer(GpuState state, int stageIndex, int index)
{
return GetTextureDescriptor(state, stageIndex, index).UnpackTextureTarget() == TextureTarget.TextureBuffer;
}
private bool QueryIsTextureRectangle(GpuState state, int stageIndex, int index)
{
var descriptor = GetTextureDescriptor(state, stageIndex, index);
TextureTarget target = descriptor.UnpackTextureTarget();
bool is2DTexture = target == TextureTarget.Texture2D ||
target == TextureTarget.Texture2DRect;
return !descriptor.UnpackTextureCoordNormalized() && is2DTexture;
}
private Image.TextureDescriptor GetTextureDescriptor(GpuState state, int stageIndex, int index)
{
return _context.Methods.TextureManager.GetGraphicsTextureDescriptor(state, stageIndex, index);
}
private int QueryInfoCommon(QueryInfoName info)
{
switch (info)
{
case QueryInfoName.MaximumViewportDimensions:
return _context.Capabilities.MaximumViewportDimensions;
case QueryInfoName.StorageBufferOffsetAlignment:
return _context.Capabilities.StorageBufferOffsetAlignment;
case QueryInfoName.SupportsNonConstantTextureOffset:
return Convert.ToInt32(_context.Capabilities.SupportsNonConstantTextureOffset);
}
return 0;
}
}
}

42
Ryujinx.Graphics.Shader/CodeGen/Glsl/Declarations.cs
View File

@ -35,23 +35,11 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
if (context.Config.Stage == ShaderStage.Geometry)
{
string inPrimitive = "points";
if ((context.Config.Flags & TranslationFlags.Unspecialized) != 0)
{
inPrimitive = DefineNames.InputTopologyName;
}
string inPrimitive = ((InputTopology)context.Config.QueryInfo(QueryInfoName.PrimitiveTopology)).ToGlslString();
context.AppendLine($"layout ({inPrimitive}) in;");
string outPrimitive = "triangle_strip";
switch (context.Config.OutputTopology)
{
case OutputTopology.LineStrip: outPrimitive = "line_strip"; break;
case OutputTopology.PointList: outPrimitive = "points"; break;
case OutputTopology.TriangleStrip: outPrimitive = "triangle_strip"; break;
}
string outPrimitive = context.Config.OutputTopology.ToGlslString();
int maxOutputVertices = context.Config.MaxOutputVertices;
@ -75,16 +63,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
if (context.Config.Stage == ShaderStage.Compute)
{
string size;
if ((context.Config.Flags & TranslationFlags.Unspecialized) != 0)
{
size = DefineNames.SharedMemorySize;
}
else
{
size = NumberFormatter.FormatInt(context.Config.Capabilities.MaximumComputeSharedMemorySize / 4);
}
string size = NumberFormatter.FormatInt(BitUtils.DivRoundUp(context.Config.QueryInfo(QueryInfoName.ComputeSharedMemorySize), 4));
context.AppendLine($"shared uint {DefaultNames.SharedMemoryName}[{size}];");
context.AppendLine();
@ -136,19 +115,12 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl
}
else
{
string localSizeX = "1";
string localSizeY = "1";
string localSizeZ = "1";
if ((context.Config.Flags & TranslationFlags.Unspecialized) != 0)
{
localSizeX = DefineNames.LocalSizeX;
localSizeY = DefineNames.LocalSizeY;
localSizeZ = DefineNames.LocalSizeZ;
}
string localSizeX = NumberFormatter.FormatInt(context.Config.QueryInfo(QueryInfoName.ComputeLocalSizeX));
string localSizeY = NumberFormatter.FormatInt(context.Config.QueryInfo(QueryInfoName.ComputeLocalSizeY));
string localSizeZ = NumberFormatter.FormatInt(context.Config.QueryInfo(QueryInfoName.ComputeLocalSizeZ));
context.AppendLine(
$"layout (" +
"layout (" +
$"local_size_x = {localSizeX}, " +
$"local_size_y = {localSizeY}, " +
$"local_size_z = {localSizeZ}) in;");

8
Ryujinx.Graphics.Shader/DefineNames.cs
View File

@ -2,14 +2,6 @@ namespace Ryujinx.Graphics.Shader
{
public static class DefineNames
{
public const string InputTopologyName = "S_INPUT_TOPOLOGY";
public const string OutQualifierPrefixName = "S_OUT_QUALIFIER";
public const string SharedMemorySize = "S_SHARED_MEMORY_SIZE";
public const string LocalSizeX = "S_LOCAL_SIZE_X";
public const string LocalSizeY = "S_LOCAL_SIZE_Y";
public const string LocalSizeZ = "S_LOCAL_SIZE_Z";
}
}

28
Ryujinx.Graphics.Shader/InputTopology.cs Normal file
View File

@ -0,0 +1,28 @@
namespace Ryujinx.Graphics.Shader
{
public enum InputTopology
{
Points,
Lines,
LinesAdjacency,
Triangles,
TrianglesAdjacency
}
static class InputTopologyExtensions
{
public static string ToGlslString(this InputTopology topology)
{
switch (topology)
{
case InputTopology.Points: return "points";
case InputTopology.Lines: return "lines";
case InputTopology.LinesAdjacency: return "lines_adjacency";
case InputTopology.Triangles: return "triangles";
case InputTopology.TrianglesAdjacency: return "triangles_adjacency";
}
return "points";
}
}
}

15
Ryujinx.Graphics.Shader/OutputTopology.cs
View File

@ -6,4 +6,19 @@ namespace Ryujinx.Graphics.Shader
LineStrip = 6,
TriangleStrip = 7
}
static class OutputTopologyExtensions
{
public static string ToGlslString(this OutputTopology topology)
{
switch (topology)
{
case OutputTopology.LineStrip: return "line_strip";
case OutputTopology.PointList: return "points";
case OutputTopology.TriangleStrip: return "triangle_strip";
}
return "points";
}
}
}

4
Ryujinx.Graphics.Shader/QueryInfoCallback.cs Normal file
View File

@ -0,0 +1,4 @@
namespace Ryujinx.Graphics.Shader
{
public delegate int QueryInfoCallback(QueryInfoName info, int index);
}

17
Ryujinx.Graphics.Shader/QueryInfoName.cs Normal file
View File

@ -0,0 +1,17 @@
namespace Ryujinx.Graphics.Shader
{
public enum QueryInfoName
{
ComputeLocalSizeX,
ComputeLocalSizeY,
ComputeLocalSizeZ,
ComputeSharedMemorySize,
IsTextureBuffer,
IsTextureRectangle,
MaximumViewportDimensions,
PrimitiveTopology,
StorageBufferOffsetAlignment,
SupportsNonConstantTextureOffset,
ViewportTransformEnable
}
}

27
Ryujinx.Graphics.Shader/ShaderCapabilities.cs
View File

@ -1,27 +0,0 @@
namespace Ryujinx.Graphics.Shader
{
public struct ShaderCapabilities
{
// Initialize with default values for Maxwell.
private static readonly ShaderCapabilities _default = new ShaderCapabilities(0x8000, 0xc000, 16, true);
public static ShaderCapabilities Default => _default;
public int MaximumViewportDimensions { get; }
public int MaximumComputeSharedMemorySize { get; }
public int StorageBufferOffsetAlignment { get; }
public bool SupportsNonConstantTextureOffset { get; }
public ShaderCapabilities(
int maximumViewportDimensions,
int maximumComputeSharedMemorySize,
int storageBufferOffsetAlignment,
bool supportsNonConstantTextureOffset)
{
MaximumViewportDimensions = maximumViewportDimensions;
MaximumComputeSharedMemorySize = maximumComputeSharedMemorySize;
StorageBufferOffsetAlignment = storageBufferOffsetAlignment;
SupportsNonConstantTextureOffset = supportsNonConstantTextureOffset;
}
}
}

105
Ryujinx.Graphics.Shader/ShaderConfig.cs
View File

@ -1,4 +1,5 @@
using Ryujinx.Graphics.Shader.Translation;
using System;
namespace Ryujinx.Graphics.Shader
{
@ -6,26 +7,100 @@ namespace Ryujinx.Graphics.Shader
{
public ShaderStage Stage { get; }
public ShaderCapabilities Capabilities { get; }
public TranslationFlags Flags { get; }
public OutputTopology OutputTopology { get; }
public int MaxOutputVertices { get; }
public OutputTopology OutputTopology { get; }
public OutputMapTarget[] OmapTargets { get; }
public bool OmapSampleMask { get; }
public bool OmapDepth { get; }
public ShaderConfig(
ShaderStage stage,
ShaderCapabilities capabilities,
TranslationFlags flags,
int maxOutputVertices,
OutputTopology outputTopology)
public TranslationFlags Flags { get; }
private QueryInfoCallback _queryInfoCallback;
public ShaderConfig(TranslationFlags flags, QueryInfoCallback queryInfoCallback)
{
Stage = stage;
Capabilities = capabilities;
Flags = flags;
MaxOutputVertices = maxOutputVertices;
OutputTopology = outputTopology;
Stage = ShaderStage.Compute;
OutputTopology = OutputTopology.PointList;
MaxOutputVertices = 0;
OmapTargets = null;
OmapSampleMask = false;
OmapDepth = false;
Flags = flags;
_queryInfoCallback = queryInfoCallback;
}
public ShaderConfig(ShaderHeader header, TranslationFlags flags, QueryInfoCallback queryInfoCallback)
{
Stage = header.Stage;
OutputTopology = header.OutputTopology;
MaxOutputVertices = header.MaxOutputVertexCount;
OmapTargets = header.OmapTargets;
OmapSampleMask = header.OmapSampleMask;
OmapDepth = header.OmapDepth;
Flags = flags;
_queryInfoCallback = queryInfoCallback;
}
public int GetDepthRegister()
{
int count = 0;
for (int index = 0; index < OmapTargets.Length; index++)
{
for (int component = 0; component < 4; component++)
{
if (OmapTargets[index].ComponentEnabled(component))
{
count++;
}
}
}
// The depth register is always two registers after the last color output.
return count + 1;
}
public bool QueryInfoBool(QueryInfoName info, int index = 0)
{
return Convert.ToBoolean(QueryInfo(info, index));
}
public int QueryInfo(QueryInfoName info, int index = 0)
{
if (_queryInfoCallback != null)
{
return _queryInfoCallback(info, index);
}
else
{
switch (info)
{
case QueryInfoName.ComputeLocalSizeX:
case QueryInfoName.ComputeLocalSizeY:
case QueryInfoName.ComputeLocalSizeZ:
return 1;
case QueryInfoName.ComputeSharedMemorySize:
return 0xc000;
case QueryInfoName.IsTextureBuffer:
return Convert.ToInt32(false);
case QueryInfoName.IsTextureRectangle:
return Convert.ToInt32(false);
case QueryInfoName.MaximumViewportDimensions:
return 0x8000;
case QueryInfoName.PrimitiveTopology:
return (int)InputTopology.Points;
case QueryInfoName.StorageBufferOffsetAlignment:
return 16;
case QueryInfoName.SupportsNonConstantTextureOffset:
return Convert.ToInt32(true);
case QueryInfoName.ViewportTransformEnable:
return Convert.ToInt32(true);
}
}
return 0;
}
}
}

22
Ryujinx.Graphics.Shader/ShaderHeader.cs
View File

@ -76,28 +76,6 @@ namespace Ryujinx.Graphics.Shader
public bool OmapSampleMask { get; }
public bool OmapDepth { get; }
public int DepthRegister
{
get
{
int count = 0;
for (int index = 0; index < OmapTargets.Length; index++)
{
for (int component = 0; component < 4; component++)
{
if (OmapTargets[index].ComponentEnabled(component))
{
count++;
}
}
}
// Depth register is always two registers after the last color output.
return count + 1;
}
}
public ShaderHeader(Span<byte> code)
{
Span<int> header = MemoryMarshal.Cast<byte, int>(code);

32
Ryujinx.Graphics.Shader/Translation/EmitterContext.cs
View File

@ -11,25 +11,15 @@ namespace Ryujinx.Graphics.Shader.Translation
public Block CurrBlock { get; set; }
public OpCode CurrOp { get; set; }
private ShaderStage _stage;
private ShaderHeader _header;
private ShaderCapabilities _capabilities;
private TranslationFlags _flags;
private ShaderConfig _config;
private List<Operation> _operations;
private Dictionary<ulong, Operand> _labels;
public EmitterContext(
ShaderStage stage,
ShaderHeader header,
ShaderCapabilities capabilities,
TranslationFlags flags)
public EmitterContext(ShaderConfig config)
{
_stage = stage;
_header = header;
_capabilities = capabilities;
_flags = flags;
_config = config;
_operations = new List<Operation>();
@ -69,24 +59,24 @@ namespace Ryujinx.Graphics.Shader.Translation
public void PrepareForReturn()
{
if (_stage == ShaderStage.Vertex)
if (_config.Stage == ShaderStage.Vertex)
{
if ((_flags & TranslationFlags.DividePosXY) != 0)
if (!_config.QueryInfoBool(QueryInfoName.ViewportTransformEnable))
{
Operand posX = Attribute(AttributeConsts.PositionX);
Operand posY = Attribute(AttributeConsts.PositionY);
this.Copy(posX, this.FPDivide(posX, ConstF(_capabilities.MaximumViewportDimensions / 2)));
this.Copy(posY, this.FPDivide(posY, ConstF(_capabilities.MaximumViewportDimensions / 2)));
this.Copy(posX, this.FPDivide(posX, ConstF(_config.QueryInfo(QueryInfoName.MaximumViewportDimensions) / 2)));
this.Copy(posY, this.FPDivide(posY, ConstF(_config.QueryInfo(QueryInfoName.MaximumViewportDimensions) / 2)));
}
}
else if (_stage == ShaderStage.Fragment)
else if (_config.Stage == ShaderStage.Fragment)
{
if (_header.OmapDepth)
if (_config.OmapDepth)
{
Operand dest = Attribute(AttributeConsts.FragmentOutputDepth);
Operand src = Register(_header.DepthRegister, RegisterType.Gpr);
Operand src = Register(_config.GetDepthRegister(), RegisterType.Gpr);
this.Copy(dest, src);
}
@ -95,7 +85,7 @@ namespace Ryujinx.Graphics.Shader.Translation
for (int attachment = 0; attachment < 8; attachment++)
{
OutputMapTarget target = _header.OmapTargets[attachment];
OutputMapTarget target = _config.OmapTargets[attachment];
for (int component = 0; component < 4; component++)
{

210
Ryujinx.Graphics.Shader/Translation/Lowering.cs
View File

@ -27,9 +27,9 @@ namespace Ryujinx.Graphics.Shader.Translation
node = RewriteGlobalAccess(node, config);
}
if (!config.Capabilities.SupportsNonConstantTextureOffset && operation.Inst == Instruction.TextureSample)
if (operation.Inst == Instruction.TextureSample)
{
node = RewriteTextureSample(node);
node = RewriteTextureSample(node, config);
}
}
}
@ -79,7 +79,7 @@ namespace Ryujinx.Graphics.Shader.Translation
sbSlot = PrependOperation(Instruction.ConditionalSelect, inRange, Const(slot), sbSlot);
}
Operand alignMask = Const(-config.Capabilities.StorageBufferOffsetAlignment);
Operand alignMask = Const(-config.QueryInfo(QueryInfoName.StorageBufferOffsetAlignment));
Operand baseAddrTrunc = PrependOperation(Instruction.BitwiseAnd, sbBaseAddrLow, Const(-64));
Operand byteOffset = PrependOperation(Instruction.Subtract, addrLow, baseAddrTrunc);
@ -124,23 +124,18 @@ namespace Ryujinx.Graphics.Shader.Translation
return node;
}
private static LinkedListNode<INode> RewriteTextureSample(LinkedListNode<INode> node)
private static LinkedListNode<INode> RewriteTextureSample(LinkedListNode<INode> node, ShaderConfig config)
{
// Technically, non-constant texture offsets are not allowed (according to the spec),
// however some GPUs does support that.
// For GPUs where it is not supported, we can replace the instruction with the following:
// For texture*Offset, we replace it by texture*, and add the offset to the P coords.
// The offset can be calculated as offset / textureSize(lod), where lod = textureQueryLod(coords).
// For texelFetchOffset, we replace it by texelFetch and add the offset to the P coords directly.
// For textureGatherOffset, we take advantage of the fact that the operation is already broken down
// to read the 4 pixels separately, and just replace it with 4 textureGather with a different offset
// for each pixel.
TextureOperation texOp = (TextureOperation)node.Value;
bool hasOffset = (texOp.Flags & TextureFlags.Offset) != 0;
bool hasOffsets = (texOp.Flags & TextureFlags.Offsets) != 0;
if (!(hasOffset || hasOffsets))
bool hasInvalidOffset = (hasOffset || hasOffsets) && !config.QueryInfoBool(QueryInfoName.SupportsNonConstantTextureOffset);
bool isRect = config.QueryInfoBool(QueryInfoName.IsTextureRectangle, texOp.Handle);
if (!(hasInvalidOffset || isRect))
{
return node;
}
@ -159,14 +154,24 @@ namespace Ryujinx.Graphics.Shader.Translation
int coordsCount = texOp.Type.GetDimensions();
int offsetsCount = coordsCount * (hasOffsets ? 4 : 1);
int offsetsCount;
if (hasOffsets)
{
offsetsCount = coordsCount * 4;
}
else if (hasOffset)
{
offsetsCount = coordsCount;
}
else
{
offsetsCount = 0;
}
Operand[] offsets = new Operand[offsetsCount];
Operand[] sources = new Operand[texOp.SourcesCount - offsetsCount];
int srcIndex = 0;
int dstIndex = 0;
int copyCount = 0;
if (isBindless || isIndexed)
@ -207,6 +212,9 @@ namespace Ryujinx.Graphics.Shader.Translation
copyCount++;
}
int srcIndex = 0;
int dstIndex = 0;
for (int index = 0; index < copyCount; index++)
{
sources[dstIndex++] = texOp.GetSource(srcIndex++);
@ -223,7 +231,9 @@ namespace Ryujinx.Graphics.Shader.Translation
offsets[index] = offset;
}
if (areAllOffsetsConstant)
hasInvalidOffset &= !areAllOffsetsConstant;
if (!(hasInvalidOffset || isRect))
{
return node;
}
@ -240,50 +250,32 @@ namespace Ryujinx.Graphics.Shader.Translation
int coordsIndex = isBindless || isIndexed ? 1 : 0;
if (intCoords)
int componentIndex = texOp.Index;
Operand Int(Operand value)
{
for (int index = 0; index < coordsCount; index++)
{
Operand source = sources[coordsIndex + index];
Operand res = Local();
Operand coordPlusOffset = Local();
node.List.AddBefore(node, new Operation(Instruction.ConvertFPToS32, res, value));
node.List.AddBefore(node, new Operation(Instruction.Add, coordPlusOffset, source, offsets[index]));
sources[coordsIndex + index] = coordPlusOffset;
}
return res;
}
else
Operand Float(Operand value)
{
Operand lod = Local();
Operand res = Local();
node.List.AddBefore(node, new TextureOperation(
Instruction.Lod,
texOp.Type,
texOp.Flags,
texOp.Handle,
1,
lod,
lodSources));
node.List.AddBefore(node, new Operation(Instruction.ConvertS32ToFP, res, value));
Operand Int(Operand value)
{
Operand res = Local();
node.List.AddBefore(node, new Operation(Instruction.ConvertFPToS32, res, value));
return res;
}
Operand Float(Operand value)
{
Operand res = Local();
node.List.AddBefore(node, new Operation(Instruction.ConvertS32ToFP, res, value));
return res;
}
return res;
}
// Emulate texture rectangle by normalizing the coordinates on the shader.
// When sampler*Rect is used, the coords are expected to the in the [0, W or H] range,
// and otherwise, it is expected to be in the [0, 1] range.
// We normalize by dividing the coords by the texture size.
if (isRect && !intCoords)
{
for (int index = 0; index < coordsCount; index++)
{
Operand coordSize = Local();
@ -292,11 +284,11 @@ namespace Ryujinx.Graphics.Shader.Translation
if (isBindless || isIndexed)
{
texSizeSources = new Operand[] { sources[0], Int(lod) };
texSizeSources = new Operand[] { sources[0], Const(0) };
}
else
{
texSizeSources = new Operand[] { Int(lod) };
texSizeSources = new Operand[] { Const(0) };
}
node.List.AddBefore(node, new TextureOperation(
@ -308,35 +300,101 @@ namespace Ryujinx.Graphics.Shader.Translation
coordSize,
texSizeSources));
Operand offset = Local();
Operand intOffset = offsets[index + (hasOffsets ? texOp.Index * coordsCount : 0)];
node.List.AddBefore(node, new Operation(Instruction.FP | Instruction.Divide, offset, Float(intOffset), Float(coordSize)));
Operand source = sources[coordsIndex + index];
Operand coordPlusOffset = Local();
Operand coordNormalized = Local();
node.List.AddBefore(node, new Operation(Instruction.FP | Instruction.Add, coordPlusOffset, source, offset));
node.List.AddBefore(node, new Operation(Instruction.FP | Instruction.Divide, coordNormalized, source, Float(coordSize)));
sources[coordsIndex + index] = coordPlusOffset;
sources[coordsIndex + index] = coordNormalized;
}
}
int componentIndex;
if (isGather && !isShadow)
// Technically, non-constant texture offsets are not allowed (according to the spec),
// however some GPUs does support that.
// For GPUs where it is not supported, we can replace the instruction with the following:
// For texture*Offset, we replace it by texture*, and add the offset to the P coords.
// The offset can be calculated as offset / textureSize(lod), where lod = textureQueryLod(coords).
// For texelFetchOffset, we replace it by texelFetch and add the offset to the P coords directly.
// For textureGatherOffset, we take advantage of the fact that the operation is already broken down
// to read the 4 pixels separately, and just replace it with 4 textureGather with a different offset
// for each pixel.
if (hasInvalidOffset)
{
Operand gatherComponent = sources[dstIndex - 1];
if (intCoords)
{
for (int index = 0; index < coordsCount; index++)
{
Operand source = sources[coordsIndex + index];
Debug.Assert(gatherComponent.Type == OperandType.Constant);
Operand coordPlusOffset = Local();
componentIndex = gatherComponent.Value;
}
else
{
componentIndex = texOp.Index;
node.List.AddBefore(node, new Operation(Instruction.Add, coordPlusOffset, source, offsets[index]));
sources[coordsIndex + index] = coordPlusOffset;
}
}
else
{
Operand lod = Local();
node.List.AddBefore(node, new TextureOperation(
Instruction.Lod,
texOp.Type,
texOp.Flags,
texOp.Handle,
1,
lod,
lodSources));
for (int index = 0; index < coordsCount; index++)
{
Operand coordSize = Local();
Operand[] texSizeSources;
if (isBindless || isIndexed)
{
texSizeSources = new Operand[] { sources[0], Int(lod) };
}
else
{
texSizeSources = new Operand[] { Int(lod) };
}
node.List.AddBefore(node, new TextureOperation(
Instruction.TextureSize,
texOp.Type,
texOp.Flags,
texOp.Handle,
index,
coordSize,
texSizeSources));
Operand offset = Local();
Operand intOffset = offsets[index + (hasOffsets ? texOp.Index * coordsCount : 0)];
node.List.AddBefore(node, new Operation(Instruction.FP | Instruction.Divide, offset, Float(intOffset), Float(coordSize)));
Operand source = sources[coordsIndex + index];
Operand coordPlusOffset = Local();
node.List.AddBefore(node, new Operation(Instruction.FP | Instruction.Add, coordPlusOffset, source, offset));
sources[coordsIndex + index] = coordPlusOffset;
}
}
if (isGather && !isShadow)
{
Operand gatherComponent = sources[dstIndex - 1];
Debug.Assert(gatherComponent.Type == OperandType.Constant);
componentIndex = gatherComponent.Value;
}
}
TextureOperation newTexOp = new TextureOperation(

2
Ryujinx.Graphics.Shader/Translation/Optimizations/GlobalToStorage.cs
View File

@ -52,7 +52,7 @@ namespace Ryujinx.Graphics.Shader.Translation.Optimizations
Operand baseAddrTrunc = Local();
Operand alignMask = Const(-config.Capabilities.StorageBufferOffsetAlignment);
Operand alignMask = Const(-config.QueryInfo(QueryInfoName.StorageBufferOffsetAlignment));
Operation andOp = new Operation(Instruction.BitwiseAnd, baseAddrTrunc, baseAddrLow, alignMask);

6
Ryujinx.Graphics.Shader/Translation/TranslationFlags.cs
View File

@ -4,9 +4,7 @@ namespace Ryujinx.Graphics.Shader.Translation
{
None = 0,
Compute = 1 << 0,
DebugMode = 1 << 1,
Unspecialized = 1 << 2,
DividePosXY = 1 << 3
Compute = 1 << 0,
DebugMode = 1 << 1
}
}

76
Ryujinx.Graphics.Shader/Translation/Translator.cs
View File

@ -16,14 +16,7 @@ namespace Ryujinx.Graphics.Shader.Translation
public static Span<byte> ExtractCode(Span<byte> code, bool compute, out int headerSize)
{
if (compute)
{
headerSize = 0;
}
else
{
headerSize = HeaderSize;
}
headerSize = compute ? 0 : HeaderSize;
Block[] cfg = Decoder.Decode(code, (ulong)headerSize);
@ -47,56 +40,21 @@ namespace Ryujinx.Graphics.Shader.Translation
return code.Slice(0, headerSize + (int)endAddress);
}
public static ShaderProgram Translate(Span<byte> code, ShaderCapabilities capabilities, TranslationFlags flags)
public static ShaderProgram Translate(Span<byte> code, QueryInfoCallback queryInfoCallback, TranslationFlags flags)
{
bool compute = (flags & TranslationFlags.Compute) != 0;
Operation[] ops = DecodeShader(code, capabilities, flags, out ShaderHeader header, out int size);
ShaderStage stage;
if (compute)
{
stage = ShaderStage.Compute;
}
else
{
stage = header.Stage;
}
int maxOutputVertexCount = 0;
OutputTopology outputTopology = OutputTopology.LineStrip;
if (!compute)
{
maxOutputVertexCount = header.MaxOutputVertexCount;
outputTopology = header.OutputTopology;
}
ShaderConfig config = new ShaderConfig(
stage,
capabilities,
flags,
maxOutputVertexCount,
outputTopology);
Operation[] ops = DecodeShader(code, queryInfoCallback, flags, out ShaderConfig config, out int size);
return Translate(ops, config, size);
}
public static ShaderProgram Translate(Span<byte> vpACode, Span<byte> vpBCode, ShaderCapabilities capabilities, TranslationFlags flags)
public static ShaderProgram Translate(Span<byte> vpACode, Span<byte> vpBCode, QueryInfoCallback queryInfoCallback, TranslationFlags flags)
{
bool debugMode = (flags & TranslationFlags.DebugMode) != 0;
Operation[] vpAOps = DecodeShader(vpACode, capabilities, flags, out _, out _);
Operation[] vpBOps = DecodeShader(vpBCode, capabilities, flags, out ShaderHeader header, out int sizeB);
ShaderConfig config = new ShaderConfig(
header.Stage,
capabilities,
flags,
header.MaxOutputVertexCount,
header.OutputTopology);
Operation[] vpAOps = DecodeShader(vpACode, queryInfoCallback, flags, out _, out _);
Operation[] vpBOps = DecodeShader(vpBCode, queryInfoCallback, flags, out ShaderConfig config, out int sizeB);
return Translate(Combine(vpAOps, vpBOps), config, sizeB);
}
@ -136,31 +94,25 @@ namespace Ryujinx.Graphics.Shader.Translation
}
private static Operation[] DecodeShader(
Span<byte> code,
ShaderCapabilities capabilities,
TranslationFlags flags,
out ShaderHeader header,
out int size)
Span<byte> code,
QueryInfoCallback queryInfoCallback,
TranslationFlags flags,
out ShaderConfig config,
out int size)
{
Block[] cfg;
EmitterContext context;
if ((flags & TranslationFlags.Compute) != 0)
{
header = null;
config = new ShaderConfig(flags, queryInfoCallback);
cfg = Decoder.Decode(code, 0);
context = new EmitterContext(ShaderStage.Compute, header, capabilities, flags);
}
else
{
header = new ShaderHeader(code);
config = new ShaderConfig(new ShaderHeader(code), flags, queryInfoCallback);
cfg = Decoder.Decode(code, HeaderSize);
context = new EmitterContext(header.Stage, header, capabilities, flags);
}
if (cfg == null)
@ -172,6 +124,8 @@ namespace Ryujinx.Graphics.Shader.Translation
return new Operation[0];
}
EmitterContext context = new EmitterContext(config);
ulong maxEndAddress = 0;
for (int blkIndex = 0; blkIndex < cfg.Length; blkIndex++)

2
Ryujinx.ShaderTools/Program.cs
View File

@ -20,7 +20,7 @@ namespace Ryujinx.ShaderTools
byte[] data = File.ReadAllBytes(args[args.Length - 1]);
string code = Translator.Translate(data, ShaderCapabilities.Default, flags).Code;
string code = Translator.Translate(data, null, flags).Code;
Console.WriteLine(code);
}