emu-mirrors/ryujinx-final
Archived
0
0
Fork 0
Code Activity

Support non-constant texture offsets on non-NVIDIA gpus

Browse source
This commit is contained in:
gdkchan 2019-12-11 03:54:18 -03:00 committed by Thog
parent 3323a3a042
commit f2c85c5d58
11 changed files with 319 additions and 33 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

13
Ryujinx.Graphics.GAL/Capabilities.cs
View file

@ -2,7 +2,8 @@ namespace Ryujinx.Graphics.GAL
{
public struct Capabilities
{
public bool SupportsAstcCompression { get; }
public bool SupportsAstcCompression { get; }
public bool SupportsNonConstantTextureOffset { get; }
public int MaximumViewportDimensions { get; }
public int MaximumComputeSharedMemorySize { get; }
@ -10,14 +11,16 @@ namespace Ryujinx.Graphics.GAL
public Capabilities(
bool supportsAstcCompression,
bool supportsNonConstantTextureOffset,
int maximumViewportDimensions,
int maximumComputeSharedMemorySize,
int storageBufferOffsetAlignment)
{
SupportsAstcCompression = supportsAstcCompression;
MaximumViewportDimensions = maximumViewportDimensions;
MaximumComputeSharedMemorySize = maximumComputeSharedMemorySize;
StorageBufferOffsetAlignment = storageBufferOffsetAlignment;
SupportsAstcCompression = supportsAstcCompression;
SupportsNonConstantTextureOffset = supportsNonConstantTextureOffset;
MaximumViewportDimensions = maximumViewportDimensions;
MaximumComputeSharedMemorySize = maximumComputeSharedMemorySize;
StorageBufferOffsetAlignment = storageBufferOffsetAlignment;
}
}
}

3
Ryujinx.Graphics.Gpu/Shader/ShaderCache.cs
View file

@ -355,7 +355,8 @@ namespace Ryujinx.Graphics.Gpu.Shader
return new ShaderCapabilities(
_context.Capabilities.MaximumViewportDimensions,
_context.Capabilities.MaximumComputeSharedMemorySize,
_context.Capabilities.StorageBufferOffsetAlignment);
_context.Capabilities.StorageBufferOffsetAlignment,
_context.Capabilities.SupportsNonConstantTextureOffset);
}
}
}

16
Ryujinx.Graphics.OpenGL/HwCapabilities.cs
View file

@ -11,11 +11,14 @@ namespace Ryujinx.Graphics.OpenGL
private static Lazy<int> _maximumComputeSharedMemorySize = new Lazy<int>(() => GetLimit(All.MaxComputeSharedMemorySize));
private static Lazy<int> _storageBufferOffsetAlignment = new Lazy<int>(() => GetLimit(All.ShaderStorageBufferOffsetAlignment));
public static bool SupportsAstcCompression => _supportsAstcCompression.Value;
private static Lazy<bool> _isNvidiaDriver = new Lazy<bool>(() => IsNvidiaDriver());
public static int MaximumViewportDimensions => _maximumViewportDimensions.Value;
public static int MaximumComputeSharedMemorySize => _maximumComputeSharedMemorySize.Value;
public static int StorageBufferOffsetAlignment => _storageBufferOffsetAlignment.Value;
public static bool SupportsAstcCompression => _supportsAstcCompression.Value;
public static bool SupportsNonConstantTextureOffset => _isNvidiaDriver.Value;
public static int MaximumViewportDimensions => _maximumViewportDimensions.Value;
public static int MaximumComputeSharedMemorySize => _maximumComputeSharedMemorySize.Value;
public static int StorageBufferOffsetAlignment => _storageBufferOffsetAlignment.Value;
private static bool HasExtension(string name)
{
@ -36,5 +39,10 @@ namespace Ryujinx.Graphics.OpenGL
{
return GL.GetInteger((GetPName)name);
}
private static bool IsNvidiaDriver()
{
return GL.GetString(StringName.Vendor).Equals("NVIDIA Corporation");
}
}
}

1
Ryujinx.Graphics.OpenGL/Renderer.cs
View file

@ -63,6 +63,7 @@ namespace Ryujinx.Graphics.OpenGL
{
return new Capabilities(
HwCapabilities.SupportsAstcCompression,
HwCapabilities.SupportsNonConstantTextureOffset,
HwCapabilities.MaximumViewportDimensions,
HwCapabilities.MaximumComputeSharedMemorySize,
HwCapabilities.StorageBufferOffsetAlignment);

3
Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGen.cs
View file

@ -133,6 +133,9 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
case Instruction.LoadStorage:
return InstGenMemory.LoadStorage(context, operation);
case Instruction.Lod:
return InstGenMemory.Lod(context, operation);
case Instruction.PackHalf2x16:
return InstGenPacking.PackHalf2x16(context, operation);

1
Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenHelper.cs
View file

@ -73,6 +73,7 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
Add(Instruction.LoadLocal, InstType.Special);
Add(Instruction.LoadShared, InstType.Special);
Add(Instruction.LoadStorage, InstType.Special);
Add(Instruction.Lod, InstType.Special);
Add(Instruction.LogarithmB2, InstType.CallUnary, "log2");
Add(Instruction.LogicalAnd, InstType.OpBinaryCom, "&&", 9);
Add(Instruction.LogicalExclusiveOr, InstType.OpBinaryCom, "^^", 10);

60
Ryujinx.Graphics.Shader/CodeGen/Glsl/Instructions/InstGenMemory.cs
View file

@ -148,6 +148,48 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
return GetStorageBufferAccessor(indexExpr, offsetExpr, context.Config.Stage);
}
public static string Lod(CodeGenContext context, AstOperation operation)
{
AstTextureOperation texOp = (AstTextureOperation)operation;
int coordsCount = texOp.Type.GetDimensions();
bool isBindless = (texOp.Flags & TextureFlags.Bindless) != 0;
bool isIndexed = (texOp.Type & SamplerType.Indexed) != 0;
string indexExpr = null;
if (isIndexed)
{
indexExpr = GetSoureExpr(context, texOp.GetSource(0), VariableType.S32);
}
string samplerName = OperandManager.GetSamplerName(context.Config.Stage, texOp, indexExpr);
int coordsIndex = isBindless || isIndexed ? 1 : 0;
string coordsExpr;
if (coordsCount > 1)
{
string[] elems = new string[coordsCount];
for (int index = 0; index < coordsCount; index++)
{
elems[index] = GetSoureExpr(context, texOp.GetSource(coordsIndex + index), VariableType.F32);
}
coordsExpr = "vec" + coordsCount + "(" + string.Join(", ", elems) + ")";
}
else
{
coordsExpr = GetSoureExpr(context, texOp.GetSource(coordsIndex), VariableType.F32);
}
return $"textureQueryLod({samplerName}, {coordsExpr}){GetMask(texOp.Index)}";
}
public static string StoreLocal(CodeGenContext context, AstOperation operation)
{
return StoreLocalOrShared(context, operation, DefaultNames.LocalMemoryName);
@ -359,17 +401,17 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
}
}
if (hasExtraCompareArg)
{
Append(Src(VariableType.F32));
}
if (hasDerivatives)
{
Append(AssembleDerivativesVector(coordsCount)); // dPdx
Append(AssembleDerivativesVector(coordsCount)); // dPdy
}
if (hasExtraCompareArg)
{
Append(Src(VariableType.F32));
}
if (isMultisample)
{
Append(Src(VariableType.S32));
@ -446,11 +488,13 @@ namespace Ryujinx.Graphics.Shader.CodeGen.Glsl.Instructions
string samplerName = OperandManager.GetSamplerName(context.Config.Stage, texOp, indexExpr);
IAstNode src0 = operation.GetSource(isBindless || isIndexed ? 1 : 0);
int lodSrcIndex = isBindless || isIndexed ? 1 : 0;
string src0Expr = GetSoureExpr(context, src0, GetSrcVarType(operation.Inst, 0));
IAstNode lod = operation.GetSource(lodSrcIndex);
return $"textureSize({samplerName}, {src0Expr}){GetMask(texOp.Index)}";
string lodExpr = GetSoureExpr(context, lod, GetSrcVarType(operation.Inst, lodSrcIndex));
return $"textureSize({samplerName}, {lodExpr}){GetMask(texOp.Index)}";
}
private static string GetStorageBufferAccessor(string slotExpr, string offsetExpr, ShaderStage stage)

1
Ryujinx.Graphics.Shader/IntermediateRepresentation/Instruction.cs
View file

@ -71,6 +71,7 @@ namespace Ryujinx.Graphics.Shader.IntermediateRepresentation
LoadLocal,
LoadShared,
LoadStorage,
Lod,
LogarithmB2,
LogicalAnd,
LogicalExclusiveOr,

23
Ryujinx.Graphics.Shader/ShaderCapabilities.cs
View file

@ -3,22 +3,25 @@ namespace Ryujinx.Graphics.Shader
public struct ShaderCapabilities
{
// Initialize with default values for Maxwell.
private static readonly ShaderCapabilities _default = new ShaderCapabilities(32768, 49152, 16);
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 int MaximumViewportDimensions { get; }
public int MaximumComputeSharedMemorySize { get; }
public int StorageBufferOffsetAlignment { get; }
public bool SupportsNonConstantTextureOffset { get; }
public ShaderCapabilities(
int maximumViewportDimensions,
int maximumComputeSharedMemorySize,
int storageBufferOffsetAlignment)
int maximumViewportDimensions,
int maximumComputeSharedMemorySize,
int storageBufferOffsetAlignment,
bool supportsNonConstantTextureOffset)
{
MaximumViewportDimensions = maximumViewportDimensions;
MaximumComputeSharedMemorySize = maximumComputeSharedMemorySize;
StorageBufferOffsetAlignment = storageBufferOffsetAlignment;
MaximumViewportDimensions = maximumViewportDimensions;
MaximumComputeSharedMemorySize = maximumComputeSharedMemorySize;
StorageBufferOffsetAlignment = storageBufferOffsetAlignment;
SupportsNonConstantTextureOffset = supportsNonConstantTextureOffset;
}
}
}

9
Ryujinx.Graphics.Shader/StructuredIr/InstructionInfo.cs
View file

@ -85,6 +85,7 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
Add(Instruction.LoadLocal, VariableType.U32, VariableType.S32);
Add(Instruction.LoadShared, VariableType.U32, VariableType.S32);
Add(Instruction.LoadStorage, VariableType.U32, VariableType.S32, VariableType.S32);
Add(Instruction.Lod, VariableType.F32);
Add(Instruction.LogarithmB2, VariableType.Scalar, VariableType.Scalar);
Add(Instruction.LogicalAnd, VariableType.Bool, VariableType.Bool, VariableType.Bool);
Add(Instruction.LogicalExclusiveOr, VariableType.Bool, VariableType.Bool, VariableType.Bool);
@ -139,9 +140,11 @@ namespace Ryujinx.Graphics.Shader.StructuredIr
{
// TODO: Return correct type depending on source index,
// that can improve the decompiler output.
if (inst == Instruction.TextureSample ||
inst == Instruction.ImageLoad ||
inst == Instruction.ImageStore)
if (
inst == Instruction.ImageLoad ||
inst == Instruction.ImageStore ||
inst == Instruction.Lod ||
inst == Instruction.TextureSample)
{
return VariableType.F32;
}

222
Ryujinx.Graphics.Shader/Translation/Lowering.cs
View file

@ -1,5 +1,6 @@
using Ryujinx.Graphics.Shader.IntermediateRepresentation;
using System.Collections.Generic;
using System.Diagnostics;
using static Ryujinx.Graphics.Shader.IntermediateRepresentation.OperandHelper;
using static Ryujinx.Graphics.Shader.Translation.GlobalMemory;
@ -23,13 +24,18 @@ namespace Ryujinx.Graphics.Shader.Translation
if (UsesGlobalMemory(operation.Inst))
{
node = LowerGlobal(node, config);
node = RewriteGlobalAccess(node, config);
}
if (!config.Capabilities.SupportsNonConstantTextureOffset && operation.Inst == Instruction.TextureSample)
{
node = RewriteTextureSample(node);
}
}
}
}
private static LinkedListNode<INode> LowerGlobal(LinkedListNode<INode> node, ShaderConfig config)
private static LinkedListNode<INode> RewriteGlobalAccess(LinkedListNode<INode> node, ShaderConfig config)
{
Operation operation = (Operation)node.Value;
@ -117,5 +123,217 @@ namespace Ryujinx.Graphics.Shader.Translation
return node;
}
private static LinkedListNode<INode> RewriteTextureSample(LinkedListNode<INode> node)
{
TextureOperation texOp = (TextureOperation)node.Value;
bool hasOffset = (texOp.Flags & TextureFlags.Offset) != 0;
bool hasOffsets = (texOp.Flags & TextureFlags.Offsets) != 0;
if (!(hasOffset || hasOffsets))
{
return node;
}
bool isBindless = (texOp.Flags & TextureFlags.Bindless) != 0;
bool isGather = (texOp.Flags & TextureFlags.Gather) != 0;
bool hasDerivatives = (texOp.Flags & TextureFlags.Derivatives) != 0;
bool hasLodBias = (texOp.Flags & TextureFlags.LodBias) != 0;
bool hasLodLevel = (texOp.Flags & TextureFlags.LodLevel) != 0;
bool isArray = (texOp.Type & SamplerType.Array) != 0;
bool isIndexed = (texOp.Type & SamplerType.Indexed) != 0;
bool isMultisample = (texOp.Type & SamplerType.Multisample) != 0;
bool isShadow = (texOp.Type & SamplerType.Shadow) != 0;
int coordsCount = texOp.Type.GetDimensions();
int offsetsCount = coordsCount * (hasOffsets ? 4 : 1);
Operand[] offsets = new Operand[offsetsCount];
Operand[] sources = new Operand[texOp.SourcesCount - offsetsCount];
int srcIndex = 0;
int dstIndex = 0;
int copyCount = 0;
if (isBindless || isIndexed)
{
copyCount++;
}
Operand[] lodSources = new Operand[copyCount + coordsCount];
for (int index = 0; index < lodSources.Length; index++)
{
lodSources[index] = texOp.GetSource(index);
}
copyCount += coordsCount;
if (isArray)
{
copyCount++;
}
if (isShadow)
{
copyCount++;
}
if (hasDerivatives)
{
copyCount += coordsCount * 2;
}
if (isMultisample)
{
copyCount++;
}
else if (hasLodLevel)
{
copyCount++;
}
for (int index = 0; index < copyCount; index++)
{
sources[dstIndex++] = texOp.GetSource(srcIndex++);
}
bool areAllOffsetsConstant = true;
for (int index = 0; index < offsetsCount; index++)
{
Operand offset = texOp.GetSource(srcIndex++);
areAllOffsetsConstant &= offset.Type == OperandType.Constant;
offsets[index] = offset;
}
if (areAllOffsetsConstant)
{
return node;
}
if (hasLodBias)
{
sources[dstIndex++] = texOp.GetSource(srcIndex++);
}
if (isGather && !isShadow)
{
sources[dstIndex++] = texOp.GetSource(srcIndex++);
}
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;
}
Operand lod = Local();
node.List.AddBefore(node, new TextureOperation(
Instruction.Lod,
texOp.Type,
texOp.Flags,
texOp.Handle,
1,
lod,
lodSources));
int coordsIndex = isBindless || isIndexed ? 1 : 0;
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;
}
int componentIndex;
if (isGather && !isShadow)
{
Operand gatherComponent = sources[dstIndex - 1];
Debug.Assert(gatherComponent.Type == OperandType.Constant);
componentIndex = gatherComponent.Value;
}
else
{
componentIndex = texOp.Index;
}
TextureOperation newTexOp = new TextureOperation(
Instruction.TextureSample,
texOp.Type,
texOp.Flags & ~(TextureFlags.Offset | TextureFlags.Offsets),
texOp.Handle,
componentIndex,
texOp.Dest,
sources);
for (int index = 0; index < texOp.SourcesCount; index++)
{
texOp.SetSource(index, null);
}
LinkedListNode<INode> oldNode = node;
node = node.List.AddBefore(node, newTexOp);
node.List.Remove(oldNode);
return node;
}
}
}